The State of Artificial Intelligence Public Affairs Worldwide

02

Opposition to AI

Opposition to AI goes beyond jobs and the electricity bill. According to Pew, in June 2026, 52% of Americans were more concerned than excited about it, up from 37% in 2021 ^[30], and 53% believe it will worsen people's ability to think creatively ^[31]. Outside the United States the mood is similar: in Ipsos's 32-country survey, AI products make 67% of Canadians and Australians nervous, along with 64% of Americans, 63% of Irish and 62% of Britons, compared with 31% of Japanese ^[32]. That opposition takes several forms. One is disdain for what AI produces: Merriam-Webster chose “slop,” low-quality content made with AI, as its 2025 word of the year ^[33]. Another is moral, held by those who consider it wrong in itself, and a third is the fear of losing abilities, which already has laboratory evidence: in an MIT experiment, people who wrote essays with a language model showed less brain activity and remembered less of what they had written ^[34]. Creators also carry the dispute over their work, which we cover in the copyright section: more than 50,000 people signed the statement calling for models not to be trained on unlicensed works ^[35].

We believe that beneath those forms lies a deeper conflict that still has no political name. AI threatens two things at once. The first is symbolic: the craft a person has mastered, the identity built on it and the sense of being useful to others. The second is material: the security of having a place in the economy. Social psychology has distinguished these two threats for decades, the realistic threat to a group's resources and the symbolic threat to its values and worldview ^[36], and studies that measured them in the face of robots and AI find both ^{[37] [38]}. That conflict almost always remains implicit. Campaigns against data centers talk about the bill, water and noise, and those who lose their jobs blame their employer. That is why we believe AI operates as an underlying cause: it is not always the stated reason, but it amplifies the reasons that are stated and adds to a discontent that comes from other processes. Because no party has articulated it, it settles on the only object of AI that a voter can reject at the ballot box: the data center.

The discontent it arrives into

AI arrives in societies that were already upset. In the first quarter of 2026, real wages were still below their 2021 level in 13 of the 37 countries the OECD measures ^[39]. The richest 1% of the world's adults held 44.5% of personal wealth in 2022, according to UBS ^[40]. The wealth of the world's billionaires, meanwhile, grew more than 16% in 2025 to a record $18.3 trillion, and for the first time there are more than 3,000 of them ^[41]. Across the 19 countries Pew surveyed, the median share who believe their children will be worse off than they are is 57%, and in France, Italy, Spain and the United Kingdom it reaches three in four or more ^[42]. In Ipsos's 30-country survey, 62% say their country's economy is in bad shape ^[43].

That discontent has already taken to the streets. In France, on September 10, 2025, 197,000 people took part in “Bloquons tout” against the austerity budget, according to the Interior Ministry ^[44]. In Madrid, on October 3, 2026, some 70,000 people according to the authorities, and half a million according to the organizers, marched over housing after Congress rejected two decrees ^[45]. In 2025 there were protests in more than 70 countries, almost always over the cost of living and corruption ^[46]. In none of them does AI appear among the demands.

We believe that does not leave it out. AI is the technology that most visibly concentrates wealth in few hands, and it is led by people who are already among the richest in the world: in October 2025 Elon Musk became the first person worth more than half a trillion dollars ^[41]. People notice. In the United States, 54% see a high risk that AI will increase inequality ^[47], and billionaires are the group most Americans want to see with less power, 62% ^[48]. Across five countries, 54% of people with low incomes believe people like them will be left behind by generative AI, compared with 31% of those with high incomes ^[49]. Where AI becomes a visible cost, there are already protests: in the United States more than 340 protests against data centers were recorded in at least 42 states, more than 80% of them in 2026 ^[50].

The threat to one's craft

Work provides more than an income. Marie Jahoda described five functions that employment serves besides the paycheck: it structures time and provides social contact, collective purpose, status and activity, and a 2023 review found that these five dimensions explain 19% of the variation in people's mental health ^[51]. That is why the harm of losing a job goes beyond money. In Germany, the non-monetary effect of unemployment on life satisfaction is much larger than the effect of losing the income ^[52], and a meta-analysis found psychological problems in 34% of the unemployed, compared with 16% of the employed, with evidence that unemployment causes them ^[53]. For 39% of American workers, their job is very or extremely important to their identity, and the share rises to 53% among those with a postgraduate degree ^[54].

That identity is what AI calls into question, because it is starting to do well the work of those who identify most with it. In OpenAI's GDPval evaluation, on tasks from 44 occupations prepared by professionals with an average of 14 years of experience, the best model of October 2025 delivered work as good as or better than the professionals' in 47.6% of cases, and in April 2026, by OpenAI's own measurement, its GPT-5.5 reached 84.9% and Anthropic's Claude Opus 4.7 80.3%, although the study itself cautions that these are precise, one-shot tasks ^[55]. The reaction has a name in the literature. Mirbabaie and his coauthors call it AI identity threat, and find that it is predicted by changes in work, loss of status and identification with one's own task ^[56]. Experienced professionals trust an algorithm's advice less than laypeople do, even when it costs them accuracy ^[57], and the same poem is valued less when the reader believes a machine wrote it ^[58]. A 2026 study puts the pieces together: people who perceive that AI can replace their occupation withdraw from their work because they lose professional identity and meaning, and gain insecurity ^[59].

Mathematicians offer an almost laboratory case. In September, an Anthropic model formalized the proof of Fermat's Last Theorem in 11 days, with 13 million lines of verifiable code ^[60], and in October OpenAI published 722 manuscripts with results the field expected to solve over decades, among them Khot's unique games conjecture, with about 42% formalized and the warning that the rest could contain errors ^[61]. Terence Tao wrote that he was horrified to see years of patient research by doctoral students and postdocs upended at a stroke ^[62]. Days later, three tenured mathematicians described on Lex Fridman's podcast what their field was going through ^[63]. None of them faced economic risk, and yet they spoke of an identity crisis among those who see themselves as problem solvers, of intellectual loneliness and of grief over each result. They recalled Andrew Wiles, who could not finish the sentence when telling that he would never again live a moment like the one at his desk the day he saw the solution, and one of them described how many academics value themselves by how others see them, which works as long as there are goals to meet ^[63]. It is what George Gurdjieff called identification: the self glued to the function one performs. That is why we believe the adjustment AI demands is psychological before it is economic, and that it is felt first by those who invested most in their craft, who are also those with the loudest public voice.

The threat is not experienced the same way everywhere. Among the general public, concern about AI is higher among those with less education ^[64], and among workers with a college degree it coexists with more hope and excitement than among the rest ^[65]. The radiologists who rejected AI at a U.S. hospital did so because of the system's opacity more than because of their identity ^[66], and professional go players played better after AlphaGo ^[67]. The threat to one's craft weighs more where the person does not see an advantage of their own in the tool.

The threat to security

The other threat touches the base of the pyramid. Abraham Maslow wrote in 1943 that, in the face of a threat, almost everything seems less important than safety, and gave as an example the preference for a job with tenure and protection ^[68]. AI touches it because it is presented in public as a substitute for people. Andy Jassy wrote in June 2025 that AI will reduce Amazon's corporate workforce ^[69], and Shopify's Tobi Lütke asked his teams to show they cannot do a job with AI before asking for more staff ^[70]. Some companies sell that replacement as their reason for being. Mechanize launched in 2025 with the goal of “the full automation of the economy” and put the wages paid by the world's employers at more than $60 trillion a year ^[71].

Every new AI capability is read twice. Whoever sees that a model can already draft a contract, read an X-ray or prove a theorem sees the benefit, and at the same time wonders whether replacing that professional was the intention from the start and whether they will be next. Psychology explains why that question weighs. The same harm is judged more severe when it is attributed to intent than when it happens by natural causes ^[72], and in the face of suffering that seems unjust, people look for an agent to blame, with corporations as the typical agent ^[73]. The statements of executives and of companies like Mechanize put the intent in plain view. According to Gallup, 27% of workers fear that technology will make their job obsolete, a record high, and among those under 45 the figure reaches 34% ^[74]. People who use AI daily or several times a week are more than twice as likely to fear their job will disappear within five years, and the fear drops when they feel their organization looks after them ^[75]. Gallup itself writes that this anxiety runs ahead of measured displacement ^[74]. That is why we believe the opposition responds more to the announced threat than to the layoff that has happened, and fear votes even if the layoff never comes.

Today the blame falls on the employer. In a Harris survey for INTOO, 69% of employees said “AI layoffs” are an excuse to cut costs ^[76], and in Gallup's only 1% of those laid off name AI as the main cause ^[77]. When the decision is presented as an algorithm's, outrage and the responsibility attributed to the company fall ^[78]. As long as the blame stays with the employer who decides, the conflict is a labor conflict. It would change in nature if someone managed to shift it toward those who make the technology.

The human in the loop

The industry answers that fear with one phrase: the human will stay in the loop. Marc Benioff said in July 2025 that the human has to be kept in the loop and that “the humans are not going away” ^[79]. Satya Nadella said in January 2024 that AI “is not about replacing the human in the loop” but about “empowering the human” ^[80]. Jensen Huang said you still want humans in the loop because they have good judgment ^[81], and in 2025 he said no one will lose their job to an AI, but to someone who uses AI, in the same answer in which he said every job will be affected ^[82]. Bill Dally, Nvidia's chief scientist, told the U.S. Senate that the way to keep control over AI is to keep a human in the loop for any really critical application ^[83]. The phrase has also reached the law: the EU AI Act requires human oversight of high-risk systems ^[84].

We believe the phrase is not enough to overcome the fear and, on the contrary, makes the risk explicit. Whoever hears it understands that their place will be to review what the machine does, which is exactly what threatens their craft. In a survey of 319 knowledge workers, work with generative AI shifted toward verifying information, integrating responses and overseeing the task, and confidence in the tool was associated with less critical thinking ^[85]. And reviewing is not the same as deciding. Automation research has known for decades that those who supervise a machine tend to over-trust it, and that this bias is not corrected by training or instructions ^[86]. In a randomized study in which 457 hospital clinicians in 13 states diagnosed simulated cases, a systematically biased AI lowered their accuracy by 11.3 points, from a baseline of 73%, and having the AI explain its prediction did not prevent the drop ^[87]. A meta-analysis of 106 experiments found that a person paired with an AI performs, on average, worse than the better of the two alone, with losses in decision tasks and gains in content-creation tasks ^[88]. The mechanism is not always over-trust: in an experiment with 227 radiologists, the doctors gave the AI less weight than it deserved, and assistance did not improve their average diagnosis even though the AI was more accurate than 78% of them ^[89]. In a smaller experiment, with 27 radiologists and a simulated AI, when the suggestion was wrong the accuracy of the least experienced fell from 79.7% to 19.8% ^[90]. The European regulation itself asks those who oversee to remain aware of the tendency to over-rely on what the system produces, known as automation bias ^[84]. Whether through too much trust or too little, the human in the loop does not guarantee a better decision, and risks merely approving what the machine has already decided.

The place in the loop is also the one that takes the blame. Madeleine Clare Elish called the position of an operator who answers for the failures of a system they barely control a moral crumple zone ^[91], and in experiments on automated cars, when the driver and the machine both err, people blame the driver more ^[92].

The phrase also suggests that the loop can shrink, and the companies themselves confirm it. A month after saying the humans were not going away, Benioff said he had reduced Salesforce's support staff from 9,000 to about 5,000 people because he needed “less heads”; according to the company, hundreds were redeployed to other areas ^[93]. Jassy wrote that AI will reduce Amazon's corporate workforce ^[69]; Sam Altman invited readers to imagine a thousand or a million virtual coworkers ^[94], and in September 2025 Anthropic's Economic Index recorded for the first time more use of AI to automate than to augment work: in the API, 77% of conversations showed automation ^[95]. Elon Musk went further in February 2026: he said that amplifying the productivity of human corporations “is simply a short-term thing” and that corporations that are purely AI and robotics will far outperform any that have humans in the loop, because a spreadsheet with some cells calculated by humans works worse than one calculated entirely by the computer ^[96]. The timelines, moreover, are getting shorter. On November 10, Anthropic told the SEC that it has met the threshold its internal policy sets for AI research accelerated by AI itself ^[97], and the AI Futures Project, author of AI 2027, a widely cited scenario about AI self-improvement, moved up to 2027 the arrival of a superhuman AI researcher, the point at which, by its definition, a company would rather fire all its human researchers and rely solely on automated ones; the group gives a 35% to 65% chance that systems able to automate all white-collar work will arrive before 2028 ^[98].

The promise reassures as long as it holds. In an experiment in Japan, telling people that humans were still in the decision loop raised trust in an AI-assisted care plan by almost 9 points ^[99]. But that assurance rests on the loop existing, and those who hear the same industry announce that it will need fewer people hear the phrase as a deadline: the human stays in the loop for now.

From fear to conflict

Faced with a threat, people and organizations tend to close up rather than adapt, an effect Staw, Sandelands and Dutton documented in individuals, groups and firms ^[100]. There are already signs at work. In a survey by Writer, a company that sells AI, 29% of employees in the United States and Europe admitted sabotaging their company's AI strategy, for example by using unapproved tools or refusing to use AI; it is self-reported and the definition is broad, but it points in the same direction ^[101]. The mathematicians on the podcast expect the field to become more guarded with its ideas, because a published idea is the piece the machine is missing ^[63]. Open conflict appears where a union channels it: SAG-AFTRA's video game performers were on strike for almost a year over the use of AI and won limits ^[11], and the Teamsters took autonomous trucks to state legislatures ^[102].

History shows what happens to a conflict like this when it has no channel. The Luddites who smashed machines in England between 1811 and 1817, in Nottinghamshire, Yorkshire and Lancashire, were not a single case, and historians disagree about how much the machine itself mattered. Eric Hobsbawm described that destruction as “collective bargaining by riot”: breaking machines was a way to pressure employers at a time when the law banned unions and most workers could not vote ^[103]. That reading fits the framework knitters of Nottinghamshire, who attacked the frames used to cut wages and turn out cheap goods. In Yorkshire, the croppers also defended their craft against the machines that eliminated it, and before that they had asked for the enforcement of old laws that banned them, repealed in 1809; in Lancashire, attacks on steam looms merged with food riots ^[104]. The dominant reading, from Hobsbawm to E. P. Thompson and Kevin Binfield, does not see Luddism as a rejection of technology in general, although Kirkpatrick Sale reads it as resistance to industrialism itself and Adrian Randall asks that it be taken as genuine opposition to new technology rather than a symptom of the price of bread ^{[103] [104] [105]}. Lord Byron said in the House of Lords in 1812 that nothing but absolute want could have driven a once honest and industrious people to such excesses ^[106]. England deployed some 12,000 soldiers against them ^[103]. In all three regions, the machine was the object on which a conflict over wages, craft and the lack of a voice was concentrated. The political scientist Murray Edelman called objects that carry emotions their subject does not explain condensation symbols ^[107].

In recent decades, workers who lost out to automation did not vote against technology. Their discontent fed parties that blamed immigration or trade ^[108], and it was driven by a sense of relative decline more than by impoverishment ^[109]. Studies find that people tend to blame immigrants and workers in other countries and to discount the effect of technology ^[110], although the Trump vote in 2016 was higher where more robots were adopted ^[111] and in Europe exposure to automation is associated with support for the radical right ^[112]. We believe AI changes that story for two reasons. Its own makers announce the substitution, so the intent no longer has to be imagined, and it has a local object that can be voted on: the data center.

The data center has the features of a condensation symbol. It leaves the costs in the community (the bill, water, noise and tax incentives) with few permanent jobs in return, while the benefit goes to companies and users elsewhere. The campaigns do not talk about AI. In the 41 ballot measures of November 3 ^[21], AI appears in 4 of the 27 campaigns we were able to document, always alongside the bill, water or the landscape. In an open-ended Gallup question to 1,561 opponents, half named resources and between 14% and 27% gave reasons specific to AI; 12% said, unprompted, that AI replaces workers ^[113]. But when it is named, AI moves the vote. In split-sample experiments, describing the same center as “for artificial intelligence” takes between 6 and 17 points off its net support ^[114]: in the Echelon survey support goes from 35 against 53 to 27 against 62 ^[114], and among Democratic voters in the ITIF survey, from 41 against 27 to 29 against 41 ^[115]. Frequent chatbot users, who therefore receive a benefit of their own, support the centers by about 20 points more ^[116]. It is what we would expect of an underlying cause: it remains implicit, but it weighs when it appears.

The weight of a minority

The figures in this section do not describe majorities, and they do not need to. Nationally, AI barely registers among concerns: in October, 1% of Americans named the advancement of computers and technology as the most important problem facing the country, the same as the environment or unemployment, compared with 17% who named the cost of living ^[117]. Where AI becomes a concrete object, the proportion changes. In a survey that compared ten types of development in the community, the data center was the only one with more opponents than supporters, 26% in favor and 46% against, below a power plant, which comes out roughly even, and below a warehouse or a wind farm, with more than 20 points in favor ^[115]. And the 27% of workers who fear technology will make their job obsolete ^[74] amounts to about 44 million people when applied to the 163.6 million employed ^[118], three times the 14.7 million union members ^[119], who are the ones winning limits on AI in their contracts today.

In public policy, intensity and a channel matter more than size. In 2021, 81% of Americans supported background checks on private gun sales ^[120]; the House passed the bill that year by 227 votes to 203 ^[121], and it did not become law. It was a broad, unmobilized majority against an organized minority. With genetically modified foods, the opposite happened. In 2016, 39% of Americans believed they were worse for one's health ^[122]. Vermont had passed a law in 2014 requiring them to be labeled from July 2016 ^[123], and that same month Congress passed a national labeling standard that replaced it ^[124]. We believe a majority does not need to reject AI for that rejection to set its course. It is enough for those who experience it as a threat to their craft and their place to find a channel for expression and amplification, and the data center already gives them one on every local ballot. That is how the tool that could be people's best ally can end up seen as their adversary.

Who articulates it

The conflict still has no owner. In the October ABC News, Washington Post and Ipsos poll, 49% trust neither party to handle AI, 26% prefer the Democrats and 19% the Republicans, while 65% see a high risk that AI will eliminate many jobs ^[47]. NBC's exit poll found the electorate split almost evenly on which party handles it better ^[125].

Some are already beginning to name it. On September 15, Bernie Sanders and Steve Bannon shared a stage in Washington to call for limits on AI, with the AFL-CIO's Liz Shuler among those attending ^[126]. Shuler had called unchecked AI “the biggest threat to workers in our lifetime” and made clear that workers are not against technology but against greed ^[127]. In October, Ruben Gallego proposed a 10% tax on data center revenue, a robot tax and a public jobs program for workers displaced by automation ^[128]. In the House, H.R. 10044 taxes model developers at a rate that rises with unemployment to fund jobs in housing, infrastructure and care ^[129]. And Abdul El-Sayed, who proposed putting half the capital of the big AI companies in public hands, won Michigan's Senate seat on November 3 ^[130].

We believe the underlying conflict will change in scale the day someone articulates it as a national issue. A party, a candidate, a union or a movement of professionals that names the threat to craft and to security would take a base that is dispersed today, and whoever articulates it will also decide whom to hold responsible. We will follow it with four indicators: candidates who make AI and jobs a national issue, bills that nationalize it, unions that move from contract clauses to political demands, and polls showing that a party is taking ownership of the issue.

The social contract

Dario Amodei wrote in January that if ordinary people lose their economic weight, the implicit social contract of democracy may stop working ^[131]. The expression has a history. Hobbes, Locke and Rousseau founded the modern state on an exchange: obedience and taxes in return for security, property and a place in the political community. For most people, that place was earned through work, and AI calls it into question.

We reviewed some 40 proposals from industry and governments to respond to that problem. The most frequent is taxing AI, robots or tokens, in about 14; next come training and worker consultation, in about 13 each; universal income or dividends, in 10; citizen or worker ownership stakes, in 9; and transition insurance, in 8. In June, Amodei proposed funding a basic income with corporate taxes, universal capital accounts and insurance that compensates those who have to accept a lower-paying job ^[132]. Anthropic's economic framework adds worker ownership stakes in AI companies ^[133]; Sam Altman proposed in 2021 a fund that distributes shares to every adult ^[134], and OpenAI, in April, a national investment fund and a formal voice for workers on the use of AI in their jobs ^[135].

What becomes law is the weakest part. In force there are notice obligations, like the European one to inform workers before using high-risk AI ^[136], and training programs. In Brazil, the rapporteur of the AI framework removed from the text the ban on mass layoffs by substitution, after pressure from industry ^[137]. In the Psiquepol + AI corpus we found 73 bills from 2025 and 2026, in 11 countries, that deal with both AI and work; almost all of them mandate studies, reports or training, and only a few, in Brazil, Italy and Mexico, propose funds, contributions or limits on substitution ^[138].

Almost no proposal addresses craft. When industry talks about meaning, it proposes detaching it from work. Amodei wrote in 2024 that meaning comes mostly from human relationships and not from economic work ^[139], and in 2026 that the link between generating economic value and each person's worth and meaning has to be broken ^[131]. Anthropic's framework says there is dignity in work, but focuses on money because it is the most tractable problem ^[133]. Money alone is not enough. In the largest guaranteed income experiment in the United States, the well-being of those who received a thousand dollars a month rose in the first year and then returned to the level of those who did not receive it ^[140]. And people prefer work to a check: in 2017, 58% supported a government program that would pay people to do tasks even if machines could do them faster ^[141]; in 2025, a survey of 12,000 people found a preference for a jobs guarantee over an income detached from work ^[142], and in 2026 another found seven times more support for an income earned through service and training than for a universal payment ^[143]. The only document in the census that puts identity at the center is the encyclical Magnifica Humanitas, which calls work a space where identity is formed and warns that economic assistance cannot be the only answer ^[144].

We believe that is the weakest point in the development of AI. The offer compensates those who lose their craft with money and shares, and nothing gives them back their craft or tells them what their place will be. The Industrial Revolution renewed its social contract decades after the conflict: factory legislation and legal unions came long after the Luddites. This time the conflict still has no owner.

And this time the window is shorter. The World Economic Forum warned in 2016 that the fourth industrial revolution heralded “the transformation of entire systems of production, management, and governance” and that the ability of governments to adapt would determine their survival ^[145]. Ten years later, Anthropic reported that AI already accelerates its own research up to the threshold set by its internal policy ^[97], and Amodei wrote in June that he worries early policy actions are at least a year out of step with AI’s progress ^[132]. No one has yet explained to the public how that transition is being thought through, and the public notices: in Pew’s February 2026 survey, 67% of Americans had little or no confidence that the government would regulate AI effectively, and 59% that companies would develop it responsibly ^[146].

We did not find a study that brings together, for artificial intelligence, the threat to identity and purpose, the threat to job security from deliberate substitution, and the escalation of the conflict from the company to politics. There is work that joins two of those pieces: identity and job insecurity in the face of AI at the individual level ^[59], symbolic and realistic threats from robots and AI ^[37], and status loss from automation with its effect on voting ^[109]. The search was broad, though not systematic.

03

The fear of job loss

Nearly eight in ten Americans believe, according to the Bentley-Gallup survey, that AI will reduce employment in the country over the next decade, six points more than a year ago, and just 5% of adults expect AI to bring more jobs twenty years from now ^{[147] [30]}. More than half also fear it will cost someone in their household a job, according to Reuters/Ipsos ^[148]. It is a conviction that cuts across parties, something rare today: in Quinnipiac, 61% of Republicans and 78% of Democrats expect fewer jobs ^[149]. On November 3, 57% of voters said AI will do more harm than good ^[125].

The fear, however, rarely points at one's own job. While 79% expect less employment overall, according to Bentley-Gallup ^[147], only 18% of workers see it as likely that the technology will eliminate theirs within five years, although the figure rose from 15% and reaches 23% at companies that have already adopted AI ^[150]. Day-to-day unease is also on the rise: according to Gallup, 27% of workers fear the technology will make their job obsolete, a record high and double the 2017 level, and among those under 45 that worry is now tied with losing benefits as their top job concern ^[74]. Gallup itself writes that this anxiety runs ahead of measured displacement, which does not lessen its political weight, because someone who fears for their children or their town votes the same as someone who has already lost their job. Young people are the most critical: according to Pew, nearly half of adults aged 18 to 29 believe AI does more harm than good, 55% say they are more worried than excited, and only one in five trusts companies to use it responsibly, while 41% do not trust them at all ^{[147] [30]}.

At the same time, those workers use it more and more. According to Gallup, 52% of employees already turn to AI in their work, 30% frequently and 15% daily ^[151]. Those who have used it say they work faster (63%) and find more creative solutions (56%), and among regular users eight in ten report that speed ^[152]. In addition, 65% of employees at organizations that implemented it say it improved their productivity ^[151]. And since 66% of those under 30 say they use an AI chatbot, whether ChatGPT, Gemini, Copilot or another ^[153], the most distrustful group turns out to be the one that uses it most. That enthusiasm is concentrated in the best jobs: college graduates use it regularly more than twice as often as everyone else (40% versus 17%), while between 70% and more than 80% of employees in health care support, production and retail sales have never touched it at work ^[152]. On top of that, 52% of workers feel they have less say than they would like over the technology their company adopts ^[152].

Many of the phrases that feed this distrust came from the industry itself. Dario Amodei, CEO of Anthropic, warned in May 2025 that AI could eliminate up to half of entry-level white-collar jobs within five years ^[154]. Jensen Huang, of Nvidia, replied publicly that he disagreed with almost all of it, and Anthropic held its position ^[154]. The firm Challenger later counted 121,401 job cuts in the United States attributed to AI between January and October, about a fifth of the total ^[155], although in October there were only 1,265, the lowest of the year, and AI fell to second place among causes, behind market conditions ^[155]. Challenger itself cautions that outside of tech and fintech these cuts are limited and that each company decides whether to cite AI, and in the Gallup survey only 1% of those laid off point to it as the main cause ^{[156] [77]}. The most cited cases also come with caveats: Klarna boasted of cutting its headcount from 5,527 to 3,098 people, and its CEO ended up admitting that customer service had worsened and that it would invest in people again; Duolingo announced it would stop using contractors for what AI could do, and its headcount grew; and Cursor says it bills more than $1 billion with about 300 people ^{[157] [158] [159]}. Those caveats rarely accompany the headline, and they leave open the suspicion that blaming AI looks better to the market than admitting weak demand.

Economists remain divided because they measure different segments. A study by Hosseini and Lichtinger of the résumés of 65 million workers found that, at companies adopting generative AI, junior employment falls by about 9% relative to those that do not, with no change at the senior level ^[160]. For its part, the Dallas Fed estimates that Texas graduates in more exposed majors are less likely to be employed and earn a lower starting salary, and the Federal Reserve Bank of St. Louis attributes to AI-linked demand about 45% of the drop in the employment-to-population ratio of young people ^{[161] [162]}. The U.S. Bureau of Labor Statistics also projects 752,100 fewer office and administrative support jobs by 2035, and names AI among the causes ^[163]. In the aggregate, by contrast, the footprint is less visible: the Yale Budget Lab finds no clear labor effect; according to the New York Fed, 61% of service firms in its district use AI and only 4% have laid off workers because of it; the Dallas Fed estimates that even if the entire adjustment were unemployment, the national rate would rise by a tenth of a point; and unemployment remained at 4.1% in October, with 165,000 new jobs and upward revisions for August and September ^{[74] [164] [165]}. Michael Barr, of the Fed, argues that AI can break career ladders even if it does not yet move the aggregate, and Lisa Cook considers it possible that graduate unemployment is partly due to this ^[166]. Skeptics point out, however, using New York Fed data, that the gap between those graduates and the overall workforce dates back to January 2021, before ChatGPT ^[167]. Anthropic's own study allows both readings, since it finds 14% fewer entries of young people aged 22 to 25 into exposed occupations, a barely significant difference, and no increase in unemployment among the exposed ^[168].

That divergence explains much of the distance between the industry and the public. The adjustment measured so far comes through hiring less, and someone who never received the offer does not appear on any layoff list, so the national rate can stay stable while recent graduates find fewer jobs. Exposure is not even either: 6.1 million workers combine high exposure with little capacity to adapt, 86% of them are women, and the largest occupations in the group are office clerks and secretaries ^[169]. The industry responds with productivity and with the jobs the technology will bring, but, according to a 2024 Pew survey cited by Stanford's AI Index, 73% of experts expect a positive effect on how people work, versus 23% of the public ^[170], and only 27% trust that companies will use AI responsibly ^[147]. We believe the fear stems from something the industry controls, because the labs sell companies the promise of doing the same work with fewer people, and as long as that is their sales pitch, the public will believe it. Daron Acemoglu and Simon Johnson argue, along these lines, that steering technology toward replacing workers, instead of expanding what they can do, is a decision made by those who develop and buy it ^[171].

Outside the United States, the World Economic Forum survey of more than a thousand employers across 55 economies sketches the same adjustment on a global scale, with 170 million jobs created and 92 million lost between 2025 and 2030, a net of 78 million, and a change of about 40% in the skills a job requires today ^[172]. That net figure, however, hides the order of events, because those who lose a job and those who fill the new ones are rarely the same people, in the same places and at the same time. Half of employers plan to reorient their business toward what AI opens up, 77% plan to train their people, 41% expect to reduce headcount where AI automates tasks, and nearly half expect to move exposed staff to other areas ^[172]. At stake is a new economy, in which work changes, companies change, and so does what each person has to know how to do to earn a living.

Behind the net figure lies a personal challenge for hundreds of millions of people. If the world's workforce were one hundred people, the Forum says, 59 would need training before 2030 and 11 would likely not get it, which amounts to more than 120 million workers at risk of losing their jobs in the medium term ^[172]. Reinventing oneself at 45, with debts and children, in another occupation and perhaps in another city, takes time, money and a support network that most people do not have, and in the United States the workers most exposed and least able to adapt are mostly women in office jobs ^[169]. Companies face a similar challenge, since the skills gap is the main barrier to transformation according to 63% of employers ^[172], and a company that reorganizes its work around AI also changes what positions it has, how it trains its people and what it sells to its customers.

That is why the greatest political and social risk lies in the transition. The first industrial revolution ended up raising wages, but in Britain, during the first half of the nineteenth century, real wages stagnated while output per worker grew and profits took an ever larger share of income, and wages caught up with productivity only after 1850, according to the economic historian Robert Allen, who called that period “Engels' pause” ^[173]. These were also the decades of Luddism, which destroyed machines, and of Chartism, which demanded political rights for workers. In the United States itself, David Autor, David Dorn, Gordon Hanson and Kaveh Majlesi found that between 2000 and 2016 the districts hardest hit by Chinese imports became polarized, with more support for extreme positions on both sides and at the expense of moderate Democrats ^[174]. An adjustment faster and more widespread than that one, concentrated in entry-level jobs and office work, can produce the same effect on a larger scale and in less time, and it would be felt first by the generation that today most distrusts AI. If the new jobs arrive a decade after the layoffs, politics is decided in that decade, and the state faces it with less revenue, because it depends largely on taxes on labor, and with more demand for support.

Dario Amodei is the one who has laid this out in the most detail from within the industry. In “The Adolescence of Technology” he argues that AI could disrupt half of entry-level white-collar jobs in one to five years, and warns that democracy's implicit social contract may stop working if most of the population loses its economic weight ^[131]. To prevent this he proposes progressive taxes, either general or targeted at AI companies, and warns against poorly designed taxes; he calls for government intervention and for displacement to be measured in real time; and he asks companies to reassign their employees rather than lay them off, to steer their customers toward innovating more than cutting costs, and to consider continuing to pay employees who no longer contribute economic value, something Anthropic says it is evaluating ^[131]. His cofounders, moreover, pledged to donate 80% of their wealth ^[131].

Universal basic income appears in this discussion as the mechanism to keep the loss of income from ending in social collapse or a fiscal crisis of the state, although the evidence is still scarce. The OpenResearch study, funded by Sam Altman, gave $1,000 a month for three years to 1,000 people in Illinois and Texas, against 2,000 who received $50, and found that they spent mostly on basic needs, worked one to two fewer hours a week, saw their individual income excluding the transfer fall by about 16 cents for every dollar received, and that the initial improvement in well-being faded ^[175]. Altman himself said in April that he no longer believes as much in basic income and prefers some form of collective ownership of compute or of equity ^[176]. That same month, Elon Musk proposed a “universal high income” paid with federal checks, without saying how it would be financed, and the libertarian magazine Reason questioned it ^[177]. Bernie Sanders introduced in June the American A.I. Sovereign Wealth Fund Act, a sovereign fund with a 50% public stake in the main AI companies and a dividend of $1,000 a year per taxpayer, which has yet to advance ^[178]. The idea also has a living precedent in a Republican state, since Alaska distributes to its residents a dividend paid with half of the distributable income of its oil fund ^[179].

We believe that, beyond the number of jobs and productivity, people are asking whether there will be a place for them in the coming economy and who pays the cost while it arrives. The Democratic Congress that takes office in January will have the Sanders proposal on the table.

China, by contrast, treats the same fear as a matter of state. Its urban youth unemployment rate, for ages 16 to 24 and excluding students, was 18% in September ^[180]. In December 2025, a Beijing labor arbitration ruled that replacing a function with AI does not justify a dismissal ^[181], and on October 23 the State Council ordered monitoring of AI-related employment risks, support for re-employment when an AI project could displace many workers, training for the displaced, and regulation of the use of AI in hiring ^[182]. These are guidelines without targets or penalties, and analysts cited by ChinaTalk, a newsletter specializing in Chinese technology policy, doubt that local governments, with tight finances, can pay for that re-employment ^[181]. We believe they matter for another reason: the same government that pushes adoption says publicly that it protects employment, while in the United States the federal government promotes AI and its response for those who lose their jobs is limited to a June re-employment guideline that treats AI as a cause of dismissal ^[12], with the rest left to the states and companies.

In the United States, the first legal response for the private sector requires disclosure, which is the cheapest thing to pass. California (SB 951, from January 2027) and Connecticut (PA 26-15, from October 2026) have already added notices of AI-related layoffs, according to the law firms Littler and DLA Piper; New York is debating a 90-day notice, and in the U.S. Senate the bill by Josh Hawley, a Republican, and Mark Warner, a Democrat, calls for a quarterly report to the Department of Labor, with a companion bill in the House, although neither has advanced ^{[183] [184] [185] [186]}. The Teamsters took jobs versus AI to Democratic governors, pushing laws that require a driver in autonomous trucks; Gavin Newsom, of California, and Jared Polis, of Colorado, have each vetoed two, and in August the union's California affiliate sued the Department of Motor Vehicles over its new rules ^[102]. None of these rules limits the business and the public supports them, since 74% believe the government does not regulate enough and 71% say they would prefer the candidate with stricter guardrails ^{[149] [187]}. With a Democratic Congress from January, the quarterly report has more room to move than before, and, if passed, it would create an official series of layoffs attributed to AI, which does not exist today and which could reduce the fear or confirm it. For its own employees, New York has already gone further, because its civil service law prohibits the use of AI systems from causing dismissal, displacement or a reduction in hours ^[188]. Other countries have older rules. Mexican law requires court authorization, absent an agreement, to reduce staff because of new machinery or procedures, with four months' wages plus twenty days per year of service; Brazil enshrines protection against automation in its Constitution; France allows dismissal for technological change only after training has been exhausted, and Spain gives workers' representatives the right to know the parameters of the algorithms that affect employment ^[189].

04

The cost of data centers

Data centers are the part of AI that a voter can see and pay for. In 2024 they consumed 192 terawatt-hours, 4.7% of U.S. electricity, and Lawrence Berkeley National Laboratory expects them to reach 11.8% by 2030, within a range of 9.5% to 15.3% ^[190]. In its 2024 report, the same laboratory estimated that their direct water consumption, mainly for cooling, went from 21.2 billion liters in 2014 to 66 billion in 2023, that hyperscale centers alone, the large campuses of the tech companies' clouds, could consume between 60 billion and 124 billion in 2028, and that generating the electricity they used in 2023 consumed nearly 800 billion liters more ^[191]. For those who believe in the race toward self-improvement, they are also the bottleneck, because without energy and permits there is no compute, and without compute there is no military or commercial advantage.

On November 3, data centers were on 41 local ballots, and in 40 of them the option limiting them won ^[21]. Ohio accounted for most of them, with three total bans and eight partial ones, while Maine approved six-month moratoriums and Kansas, California and Florida, bans ^[21]. In Wells, Maine, more than 80% voted for a retroactive moratorium ^[192]. The only exception was Harwood, North Dakota, where a $3 billion project is already under construction ^[21]. The polls pointed in the same direction: according to Pew, between January and August the share of the public that considers data centers mostly bad for household energy costs went from 38% to 50%, and 76% of voters oppose one in their community ^{[193] [125]}. That rejection coexists, however, with use. In May, 65% of adults used AI every week, versus 52% in February, and in October 72% opposed a data center in their community, 66% opposed building more in the country and 57% said AI is bad for society ^{[194] [47]}. Neither the Edison Research and SSRS survey nor the ABC News, The Washington Post and Ipsos poll cross-tabulates users with opponents.

The bill is the opposition's central argument. The Independent Market Monitor for PJM, the grid that runs from Illinois to Virginia, attributes to data centers $29.369 billion in costs across four capacity auctions ^[28], and Heatmap, an outlet specializing in climate and energy, counts at least $260 billion in projects canceled in 2026 due to local opposition ^[195]. The bill figure has a counterweight in so-called “phantom load”: according to the consultancy Wood Mackenzie, grid operators and utilities will commit to about 28% of the 1,066 gigawatts data centers request, and planning the grid for the rest makes service more expensive for everyone ^[196]. The opposition also extends to air and water. On October 30, a federal court in Mississippi refused to dismiss the NAACP lawsuit against xAI over the Southaven turbines, which according to the environmental organization Earthjustice operate by the dozens without a permit, and had earlier rejected a similar motion by the Department of Justice, which intervened in favor of the company ^[197]. Water also reached the courts in the case of Amazon. In July 2026, Nathan Wangusi, manager of AWS's water sustainability program until September 2024, sued the company in Arlington County, Virginia, because, according to water utility bills he obtained through public records requests, the 42% reduction it announced in the northern part of the state does not match what was billed. AWS responds that an outside provider, which it does not name, verified its 2025 data, and for now these are allegations without a ruling. In March, Amazon had agreed to pay $20.5 million to residents of northeastern Oregon over nitrates in their wells, without admitting fault ^[198].

Since November 3, the central decisions have shifted to state governors. Texas already had a statewide pause, because in August Abbott ordered its regulator and ERCOT, the operator of the Texas power grid, to suspend new connections pending an audit, and ERCOT postponed the study of about 200 eligible gigawatts, of which it expects 65 to be built by 2032 ^[22]. Only 17% of data centers answered the mandatory water survey ^[22]. That state is home to Stargate Frontier, of 1,820 megawatts, and Nexus Hubbard, of 6,400 megawatts, mostly for Anthropic ^[199]. Hinojosa won with 57% and promises an immediate moratorium, an end to the tax exemptions for large data centers, which she says cost more than $1 billion a year, and a tax on the electricity they consume; however, in 2023 she voted to expand those exemptions, and everything she promises depends on a legislature where Democrats have just won the state House for the first time in 24 years ^{[200] [201] [202] [203] [199]}. In Ohio, Amy Acton proposes a moratorium conditioned on data centers paying for their gas, water and electricity and proving that they do not raise anyone else's rates, on an end to confidentiality agreements with local officials, on building, maintaining and operating with union labor, on complying with water and air standards, on locating on abandoned industrial land, and on returning incentives if they do not keep their job promises ^[24]. In Georgia, Keisha Lance Bottoms proposes a statewide pause to study the impact, and the commission that regulates electricity rates will have a Democratic majority for the first time in more than twenty years ^{[200] [25]}. Shapiro, reelected in Pennsylvania, made his standards mandatory in August; they call for on-site energy paid for by the developer, infrastructure costs covered even if the center closes, these centers being first in line for cuts during an emergency, and, if the community rejects the project, the state rejecting it too ^[204]. In New York, Kathy Hochul decreed a one-year moratorium in July ^[205].

Since December 2025, 28 bills with data centers in the title have been introduced in the U.S. Congress, and 21 of them were filed from June 24 onward, so that in one summer the issue went from local news to a congressional docket, with 23 Democratic bills and 4 Republican ones ^[206]. The voluntary route was exhausted before the election. H.R. 9340, which asks data centers to cover their costs without requiring them to, passed the House 417-3 and failed twice in the Senate, the last time on September 30, with 57 of the 60 votes it needed, and Chuck Schumer dismissed it for having no teeth ^{[207] [208]}. The permitting bill from Shelley Moore Capito, Mike Lee, Sheldon Whitehouse and Martin Heinrich, from both parties, does impose obligations: any load of 20 megawatts or more would pay the cost of the grid built for it, with a guarantee before construction and with payment still due even if the center leaves ^[209]. As of this writing, it shows no progress since the election. The payment pledge that the White House had the tech companies sign was expanded in July to more than 200 signatories, which according to the White House itself cover 80% of the energy delivered, but it still lacks an enforcement mechanism ^{[210] [211]}. Microsoft in January and Anthropic in February had already promised on their own not to pass the increase on to households ^[212]. Bernie Sanders renewed his call for the national moratorium he announced in March with Alexandria Ocasio-Cortez, introduced as S. 4214 in the Senate and H.R. 9442 in the House since June ^[9], and Yassamin Ansari and Greg Casar call for action from day one ^{[213] [214]}. Hakeem Jeffries, by contrast, said before the election that such a pause was not a position he had put forward, John Fetterman calls it a gift to China, and Trump retains the veto ^{[215] [216]}.

National security opens routes that bypass the county. Executive Order 14318, of July 2025, directed agencies to accelerate permits for projects that invest at least $500 million, add more than 100 megawatts of load or are designated for national security, in addition to offering federal land and allowing the Defense Department to lease land on its installations, and it revoked Biden's order of January of that year on the same subject ^[217]. The Department of Energy chose four of its sites, in Idaho, Oak Ridge, Paducah and Savannah River ^[218], and last July it selected Brookfield and NextEra for Paducah, with about $100 billion announced and on-site gas and battery generation ^[219]. The Army chose Carlyle at Fort Bliss, Texas, and CyrusOne at Dugway, Utah, in March, and presented AI as a strategic asset ^[220], while the Air Force offered five bases and at Arnold received no bids ^[221]. In April, Trump invoked the Defense Production Act for transformers and other grid equipment ^[222], an authority that expires on December 11, 2026, if Congress does not renew it ^[223]. The other way out is to build outside the country. The first phase of Stargate in Abu Dhabi, 200 megawatts within a 5-gigawatt campus, was scheduled for the third quarter of this year, and the Commerce Department authorized in November 2025 the sale of advanced chips to Humain, of Saudi Arabia, and to G42, of the Emirates ^[224].

Those routes do not bypass the grid or state rates. Federal land avoids county zoning, but the energy still arrives through a state's grid and at rates approved by a state regulator. In addition, the December 2025 order seeking a federal framework for AI directs that the legislative proposal not preempt state rules on compute and data center infrastructure, other than “generally applicable permitting reforms” ^[225]. That carve-out applies to the legislative proposal, while the litigation task force created by the same order is left without that express limitation. In the House, meanwhile, there is already a bill to protect local control over them ^[226], and we found no federal preemption in force, that is, no federal rule that overrides local permits. France resolves it by law, since a decree can declare a data center a project of major national interest. Germany goes in the opposite direction and requires those operating from July 2026 to reach an efficiency of 1.2, measured as the center's total energy divided by what its equipment consumes, and to reuse at least 10% of their energy ^[227].

Building abroad moves the problem to another jurisdiction. On October 28 the United Kingdom published its Data Centre Compact, which calls for credible commitments on the grid capacity requested, an energy efficiency of 1.3 or better, annual energy and water reports and, for strategic centers, a local benefit plan, and in exchange offers to reserve and reallocate grid connections and accelerate them before 2030 ^[228]. In Europe, more than $42 billion in projects have been delayed, modified or canceled due to public concern, the Danish grid operator paused new large-scale contracts in the face of requests for about 60 gigawatts, and Johor, in Malaysia, stopped approving Tier 1 and Tier 2 centers ^[27]. The Gulf has little local opposition, but carries another risk: on March 7, Iranian drones damaged three AWS data centers in the Emirates and Bahrain ^[229]. The most distant way out is orbit. SpaceX asked the FCC, the federal telecommunications regulator, for authorization for up to one million computing satellites, and Alphabet said in October that it has already launched its TPUs, the AI chips designed by Google itself, into space with Project Suncatcher ^{[230] [231]}. That route bypasses the county, but falls under the FCC, which opened the docket for comment and which the Secure World Foundation, an organization dedicated to the sustainable use of space, is asking to deny or defer the waivers until there is a complete technical record on debris, congestion and environmental impact ^[230]. For now, what has been published consists of applications and prototypes.

Demand keeps growing while construction slows. Alphabet's model APIs, the connections through which developers use its models, process nearly 32 billion tokens per minute, 40% more than the previous quarter, and the company warned that it is starting to use third-party compute to meet that demand, with pressure on its margins ^[231]. Michael Cembalest, of JPMorgan, estimates that most of the 2027 construction will not be affected, although 37.5 gigawatts planned for 2028 could be paused and another 15 made subject to conditions ^[232]. The stock market separated winners and losers the day after the election. Microsoft rose 6.7%, Amazon 6.1% and Alphabet 3.3%, while Oracle, construction contractors, GE Vernova and the specialized clouds fell; the latter are companies that rent compute for AI, which took on debt to build and pay interest while revenue is delayed ^[232].

We believe the slowdown favors those who already have data centers in operation and the cash to wait. A delay weighs little on Microsoft, but it can bankrupt those who owe money and promised a delivery date, and in that case their sites, equipment and contracts would end up in the hands of the big players. A startup would depend more on the compute of those who compete with it, and those who voted to slow data centers could end up with fewer competitors facing the big companies, which will give material to the antitrust debate. Value shifts to the specific site: energy with a firm date and a signed local agreement weigh more than a low-priced plot of land, as the Harwood exception suggests, and a community willing to accept a data center gains power to negotiate costs, jobs and investment. If construction leaves the country, moreover, it takes with it the construction jobs that today sustain union support, and the service is sold back to American customers, so the county is left without the construction and with the same AI.

Energy is the other constraint, and no source solves it alone within the timeframe of the race. Gas is the fastest to connect and therefore dominates on-site generation: Louisiana approved three Entergy plants, about 2.26 gigawatts, for Meta's Hyperion data center, and Mississippi allowed xAI 41 turbines in Southaven, a decision that is under appeal ^[23]. Turbines, however, are scarce, and GE Vernova has accumulated 122 gigawatts in orders and reservations, expects to end the year with at least 130, is already committing deliveries for 2031, and about a fifth of its customers are data centers ^[29]. Solar and batteries dominate new capacity: of the 86 gigawatts expected in 2026, 51% are solar and 28% batteries, versus 7% gas ^[233], although the 2025 tax law requires wind and solar projects that begin construction after July 4, 2026, to enter service before the end of 2027 to keep their credits ^[234]. Nuclear is the long-term bet of the government and the industry. Trump set the goal of quadrupling capacity, to 400 gigawatts by 2050 ^[235]; Meta contracted up to 6.6 gigawatts with Vistra, Oklo and TerraPower through 2035, Google 500 megawatts with Kairos, and Microsoft expects to reopen Three Mile Island, now Crane, in 2027, although PJM speaks of transmission upgrades through 2031 ^[236]. Geothermal and fusion, with Google's agreements with Fervo and Microsoft's with Helion, are still measured in tens of megawatts ^[237].

The lowest-cost option is to make better use of the existing grid. A Duke University study estimates that 76 gigawatts of new load would fit if data centers agreed to curtail 0.25% of their annual consumption, about 85 hours, and 126 gigawatts if they curtailed 1% ^[238]. Along these lines, Google signed agreements in August 2025 with Indiana Michigan Power and the Tennessee Valley Authority to reduce machine learning loads when the grid is under stress, although it cautions that this flexibility exists only at some sites. On September 22, the same company withdrew, without giving reasons, its petition for a data center of at least $1 billion in Franklin Township, Indianapolis, ahead of a vote in which at least 17 council members had announced their opposition ^[239]. In 2026 the Department of Energy authorized PJM and ERCOT to use data centers' backup generators as a last resort ^[240], and in June FERC, the federal energy regulator, opened proceedings against each regional operator to set how large loads connect ^[241]. On October 9, FERC rejected the core of PJM's proposal, that is, the connection of large loads with priority curtailment and the obligation to bring their own capacity, for encroaching on the states' ratemaking jurisdiction, and left curtailment rules to the states ^[242].

The industry defends data centers with construction jobs, local revenue and rates that protect households. Construction unions pack hearings in favor, because at one Columbus trades council data centers account for at least 40% of hours worked, and at IBEW Local 26, the electricians' union, at least 50% ^[243]. The AFL-CIO, the country's largest labor federation, opposes moratoriums and calls for regulating them ^[243]. Even so, the labor movement does not move in a single direction: in July, National Nurses United and other organizations called in North Carolina for a moratorium on new data centers until the law protects patients and workers from AI ^[244]. Virginia's legislative audit commission estimates 74,000 jobs in the state and up to 31% of local government revenue where the industry is mature, with most of the benefit coming from construction ^[245]. A survey paid for by the American Infrastructure Alliance found that 58% oppose data centers in their state, but that 54% would accept them if they pay for their electricity and protect rates and the environment, while 31% would oppose them anyway and seven in ten fear that a moratorium would give China an advantage; the release does not publish the sample or dates ^[246]. Rates with minimum payments already work. In Virginia, from January 2027, customers of 25 megawatts or more will pay at least 85% of the transmission and distribution cost they reserve, and Ohio, Indiana, Oregon, Minnesota and Wisconsin have similar arrangements ^{[247] [248]}. Several states, moreover, have already written it into law. Florida requires each customer of 50 megawatts or more to cover its full cost of service, and on November 3 its commission rejected the first rate filed under that law, Duke Energy Florida's, for failing to prove that it covered that cost ^[249]; Idaho subjects new loads of that size to a “no harm” test, South Dakota charges the data center the cost of serving it even if it leaves the system, Delaware orders it to be cut off before other noncritical loads, and Texas requires new loads to be able to disconnect during a firm load shed ^[250]. Dominique Toublan, of Barclays, describes the opposition as bipartisan and local, and believes it makes construction more expensive without stopping it ^[251].

China does not face that negotiation. In 2025 it added 542 gigawatts of electric capacity without putting it to a vote in any county ^[252]. That year it added about 315 gigawatts of solar and 119 of wind, but also about 78 of coal, the highest in the series of the Centre for Research on Energy and Clean Air (CREA), an independent research center, and it has 36 reactors under construction, nearly half of those in the world ^[253]. The United States, for its part, consumed about 45% of the world's data center electricity in 2024, and the International Energy Agency expects that demand to grow 130% by 2030 and to account for nearly half of the increase in the country's electricity demand ^[254]. China's “AI+” plan orders making use of the national nodes of the “East Data, West Computing” program, which moves compute to regions with cheaper energy and land ^[255]. That model leaves idle capacity. In 2023 and 2024 more than 500 projects were announced, according to KZ Consulting, many pushed by local governments without demand to justify them, and in early 2025 the Chinese outlets Jiazi Guangnian and 36Kr reported that up to 80% of the new compute remained unused ^[256]. The United States builds more slowly because each county can say no, while China builds fast and sometimes too much, and asking counties to vote differently does not change the former.

The moratorium already has a text outside the United States. Andrew Gee introduced it in Canberra on September 7 together with mandatory shutdown, François Ruffin filed in Paris, on October 13, a motion for a resolution against non-sovereign data centers, and Die Linke brought its motion to the Bundestag ^[257]. Korea and Brazil, by contrast, legislated in the opposite direction: Korea enacted a special law for AI data centers, and Brazil approved Redata, a tax incentive regime for data centers, with a partial veto in September, while its general AI framework awaits committee report in the Chamber and there are already bills to charge them for water and limit their use of it ^[26]. We believe the investment argument can no longer be taken as won in any parliament.

In our view, this front will no longer be closed with voluntary promises. Voters chose to impose limits on 40 of 41 ballots, and four of the seven states where the American Infrastructure Alliance wanted to negotiate standards in 2027 elected governors who promise a moratorium or conditions ^[258]. Acton's conditions (paying the full cost, transparency with local authorities and union labor) are the most likely template for that negotiation, and parts of that list are already law: Delaware prohibits its counties from signing confidentiality agreements with data centers of 100 megawatts or more, New Jersey allows project labor agreements and Virginia requires a site assessment before rezoning ^[259]. The national security argument opened up federal land, equipment and permits ^[217], and we believe it clashes with what local communities voted for.

05

Many users, few defenders

AI has millions of users and few organized defenders. Nearly half of American adults (49%) say they use an AI chatbot, and among those under 30 the figure reaches 66%; ChatGPT is the most used (44%), followed by Gemini (24%), Copilot (17%), Meta AI (14%), Grok (8%) and Claude (6%) ^[260]. But use did not turn into votes, since barely 8% used a chatbot to get information about the election ^[261]. The economic base, moreover, is thin. A $10 billion to $20 billion data center leaves 500 to 1,250 permanent jobs, and the labs' offices follow talent, so they are concentrated in a few cities ^[262]. There is no national network of permanent employment comparable to the 1,800 suppliers in 48 states that the F-35 program reports ^[263]. The base that does exist is on the construction site: more than 216,000 construction jobs tied to data centers since 2022, five unions in the American Infrastructure Alliance, and training agreements with the trades, the largest of which is the one Meta signed in August with the apprenticeship programs of the 14 building trades unions, which have 3.2 million members combined, with no amount or targets published ^{[264] [258] [265]}. Cloud providers and chipmakers, by contrast, do have plants and operations in many states ^[266].

Industry money bought few victories. Leading the Future, the super PAC (that is, a committee that can raise and spend without limit on campaigns as long as it does not coordinate with candidates) funded by AI founders and investors, with the venture capital firm Andreessen Horowitz and Greg Brockman, president of OpenAI, among its largest donors, raised $76 million, of which it spent about $26 million in primaries and less than $10 million in the general election ^{[267] [268] [269]}. Its federal committee reported $23.8 million in cash on hand as of October 14, and Fortune attributes $34 million to its entire network as of October 26, after Brockman withdrew a second contribution of $25 million ^{[268] [269]}. In the primaries it won its most expensive fight, in which it spent more than $8.1 million against Alex Bores, who lost in NY-12 even though Public First, a super PAC that calls for more regulation, put between $12 million and $15 million behind him ^[270]. In the general election, however, its Republican candidates fell with the Democratic wave. It put $4.8 million behind Darline Graham in the South Carolina Senate, and the winner was Annie Andrews, who nicknamed her “Data Center Darline”; its candidates in Montana and North Carolina also lost ^[271]. Public First, which spent about $28 million in the cycle, won with its Democratic candidates in Ohio and Pennsylvania and lost almost everything with its Republicans ^{[271] [268]}. Scott Wiener, author of SB 1047, the AI safety bill that Newsom vetoed in 2024, won Nancy Pelosi's seat with 54% and the support of Public First ^[272]. Anthropic contributed $40 million to Public First, restricted to non-electoral activity ^[273]. Neither result isolates the effect of money. Meta withdrew $25 million of the $65 million it had committed to four state super PACs ^[274]. Neither party owns the issue: in the September AP-NORC poll, 64% said AI is advancing too fast, including 76% of Democrats and 54% of Republicans, and 39% trust neither party to handle it ^[275].

The most useful comparison is the defense industry, because of the way it secured defenders in almost every part of the country for military policy in general, that is, how much is spent and when force is used, rather than for any particular aircraft. Eisenhower warned of it in his 1961 farewell, when he urged guarding against the unwarranted influence of the military-industrial complex ^[276]. That distribution began as a government decision. In World War II, the Plant Site Board, the federal board that chose where to locate plants, took them to rural areas with no other industry, even though companies preferred cities, and the Cold War kept them there ^[277]. Today the Pentagon spends in all 50 states: $606.7 billion in fiscal year 2024, of which $423.3 billion was in contracts ^[278]. In Hawaii and Virginia that spending equals 10% of state output, and in Connecticut, Alabama, Maine and Mississippi between 4% and 5.5%; in addition, there are counties tied to a single program, such as Tarrant, Texas, to the F-35, Pima, Arizona, to Raytheon's missiles, and New London, Connecticut, to Electric Boat's submarines ^[278].

Final assembly plants are split between the two parties. Many are in rural Republican districts: Lockheed integrates missiles in Troy, Alabama, and in Camden, Arkansas; General Dynamics makes Abrams tanks in Lima, Ohio; Northrop, rocket motors in northern Utah; and Huntington Ingalls, destroyers in Pascagoula, Mississippi. Others are in Democratic districts, such as the submarines in Groton, Connecticut, and Raytheon's radars in Andover, Massachusetts ^[279]. The supplier chain is spread on top of them, which is what critics call political engineering. Lockheed says it has suppliers in every state and almost 1,900 for the F-35 alone, and Northrop about 400 in 40 states for the B-21 bomber ^[280]. The practice is old: Rockwell spread more than 5,000 B-1 subcontracts across 48 states, and when Congress sought to cut the B-2 in the 1980s, Northrop showed work in 46 states and in at least 383 of the 435 districts ^[281].

The effect shows up in overall policy more than in each vote. Almost every study found little relationship between a district's contracts and its representative's vote, and Kenneth Mayer concluded that spreading subcontracts changes little about how much is spent and what is bought ^[282]. Rebecca Thorpe explains the difference: in cities with diverse economies the military contract carries little weight, but representatives from rural districts that depend on it provided the margin to maintain a large military. That is why there was no demobilization after Korea, Vietnam or the Cold War, and those same representatives more often supported the wars of presidents from the opposing party, so that Congress stopped restraining the Executive ^[277]. Larry Bartels calculated that district opinion added about $17 billion, close to 10%, to the 1982 defense budget, and that Pentagon spending in each state also pushed it up ^[283].

The result is a budget that grows without the majority asking for it. In February 2024, 35% of Americans said too much is spent on defense, 33% that the amount is about right and 29% that too little is spent ^[284]. Even so, Congress passed the fiscal year 2026 defense bill 312-112 in the House and 77-20 in the Senate, which, according to law firms such as Holland & Knight, extends the unbroken streak to 65 years ^[285], and spending reached $1.009 trillion with what the reconciliation law added, while the request for 2027 is $1.45 trillion ^[286].

The base does not save every program. Congress funded tanks the Army did not request in order to sustain the only plant that built them, and the F-35 caucus cites jobs by state to defend the program ^{[287] [263]}. But the F-22 had suppliers in 44 states and was canceled 58-40, and the Republicans who signed letters for the climate law's plants later voted for the 215-214 cut, which protected only what had already begun ^{[288] [289]}. Something similar happens outside defense: TikTok, with an enormous user base, lost in the House 352-65, and the pharmaceutical industry, with all its prescribing physicians, lost the Medicare price negotiation in 2022 ^[290]. In defense, by contrast, the buyer sits in Congress, and that is why the base protects the level of spending more than each weapon.

A base like that could form through the industry's geography or through the military budget, and both paths have limits. The industry's geography does not resemble that of defense: a data center leaves few permanent jobs, so the dependence it creates is fiscal, up to 31% of local government revenue in Virginia where the industry is mature ^[245], and even so local measures went against it on 40 of 41 ballots ^[21]. The other path is to enter the military budget, which is already the defense with the most weight. The Pentagon contracted the four large labs ^{[291] [292]}, the White House opened federal land to data centers on the argument of the race ^[217], and the fiscal year 2027 request seeks $58.5 billion for AI, of which $46 billion is for a “sovereign AI arsenal” ^[286]. Seven in ten respondents in the American Infrastructure Alliance survey also fear that a moratorium would give China an advantage ^[246]. New defense companies are already copying the old model, such as Anduril, which is building a plant in Ohio with 4,008 committed jobs and $310 million in state support ^[293]. Critics see there the risk Eisenhower pointed out. William Hartung and Michael Brenes, of the Quincy Institute, a Washington think tank that promotes military restraint, warn that venture-backed companies aspire to be the new military contractors ^[294], and Roberto González, of the Costs of War project at Brown University, calculates that the Pentagon's five largest contracts with big tech companies added up to ceilings of at least $53 billion between 2019 and 2022 ^[295].

The spending that did grow is lobbying. In the third quarter of 2026, Nvidia reported $2.15 million, Anthropic $2.11 million and OpenAI $1.53 million, records for each ^[296], although they remain below Meta, at $5.88 million, Amazon, at $4.49 million, and Google, at $3.23 million ^[297]. We believe, however, that these figures are the weakest indicator of a company's influence in government, because they measure what the law requires to be reported. Real weight forms in what no form captures, such as direct access to decision makers, officials who left the company or joined it, seats on advisory boards, contracts that make an agency dependent on a vendor, and the think tanks that write the first draft of a proposal. It also forms in the districts, where the legislator listens to whoever creates the jobs, pays local taxes and buys from suppliers in his or her electorate. Yaqiu Wang and Nicholas Bequelin observe in Project Syndicate that almost every proposal to limit AI companies receives the same response, that it would give China an advantage, and they argue that this defense weakens what the United States does have over China, which is an open society with democratic oversight ^[298].

In China, AI does not need a base to defend it, because the State defends it, having merged the civilian and the military by decree. In 2017 it created the Central Commission for Integrated Military and Civilian Development, chaired by Xi Jinping, and its AI plan of that year mandates military-civil fusion in AI; the Pentagon notes that the People's Liberation Army can commission research from the private sector through that channel ^[299]. Local governments also compete for it. Beijing, Shanghai, Shenzhen and other cities hand out compute vouchers to startups, generally of $140,000 to $200,000, and Zhejiang, the province of DeepSeek, offered to cover up to 8 million yuan; those vouchers pay for demand on the data center network the State is building under “East Data, West Computing” ^[300]. No one votes on those subsidies or can stop them, and the same model left data centers unused ^[256].

We believe today's base is temporary, divided and concentrated on the construction site, with few permanent jobs and no voices in the trades that feel affected. Defense shows what a well-distributed base is worth: it protects the level of spending and the Executive's freedom to use force, though not each program, and it was built by the State when it decided where to produce. AI does not have that geography, and the shortcut of becoming part of the military budget would give it defenders in Congress at the cost of falling under the same suspicion Eisenhower named and of depending on a single customer. Direct electoral spending also showed its limits, since a candidate can lose for being associated with data centers and no one won by defending them. The United States, for its part, cannot copy the Chinese model. We have not measured the closest civilian analog, which is the platforms that mobilized users and suppliers against local regulators, nor exposure by district, which has no public measurement.

06

The economic promise

The economic promise of AI is enormous, though it has yet to be proven. In April 2023, Goldman Sachs estimated that generative AI could raise global GDP by 7%, nearly $7 trillion, and add 1.5 points to annual productivity growth over ten years ^[301]. McKinsey estimated in June of that year that it could contribute between $2.6 trillion and $4.4 trillion a year across the 63 use cases it analyzed, and between 0.1 and 0.6 points a year to labor productivity through 2040 ^[302]. Daron Acemoglu, of MIT, arrives instead at an order of magnitude less, estimating that AI would raise total factor productivity, the part of growth not explained by more capital or more labor, by at most 0.66% over ten years, and by less than 0.53% if hard-to-learn tasks are discounted ^[303].

Field evidence from 2026 points to gains in programming, though not yet in companies' results. METR, which in early 2025 had measured that programmers using AI took 19% longer, repeated the experiment from August 2025 with agents such as Claude Code and Codex, 57 programmers and more than 800 tasks: with AI, tasks took between 4% and 18% less time. The confidence intervals include zero and METR considers the data a weak signal, though probably below the real effect, because many programmers refused to work without AI and left out the tasks they did not want to do without it ^[304]. At Microsoft, among tens of thousands of engineers, those who adopted Claude Code or GitHub Copilot CLI in early 2026 merged about 24% more pull requests than they would have without those tools ^[305]. Meanwhile, in a summer 2026 survey of almost 9,000 executives at firms in seven countries, 78% of firms use AI, but over three years their net employment did not change and their labor productivity growth rose by only 0.2% a year; they expect 0.6% a year over the next three ^[306].

Investment, by contrast, is booming. In the United States, spending on information processing equipment grew 37.4% annualized in the third quarter ^[307]. The big tech companies plan to spend more than $740 billion this year, compared with about $400 billion the previous year, and Goldman Sachs estimates $7.6 trillion in AI-linked investment between 2026 and 2031 ^{[308] [251]}. Our calculation with official data attributes to compute, software and data centers 1.03 of the 2.08 points of growth over the last four quarters ^[309]. That argument turns AI into economic policy, and Trump used it after the stock market drop, when he blamed Democratic voters and spoke of the losses in retirement accounts ^[310].

Economists, however, do not agree on that figure. Computer imports subtract 1.32 points over the same period, because much of the equipment is manufactured abroad, and net trade subtracted 3.3 points in the third quarter ^{[309] [307]}. Jan Hatzius, of Goldman Sachs, argued in February that AI investment contributed almost nothing to 2025 GDP, precisely because of imports ^[311]. The St. Louis Fed attributed to AI categories 39% of growth in the first nine months of 2025, versus 28% for information technology in 2000, without clarifying how it treats imports ^[312]. ING, with an import-net measure, puts it at 36% in the second quarter of 2026, while Jason Furman, a Harvard economist, reached 92% for the first half of 2025 by adding up all equipment and software ^[313]. All these figures are accounting exercises and none measures what would happen without that investment, whose size Goldman estimates at about $600 billion, close to 2% of GDP ^[314].

Citizens, for their part, see little of the growth the industry claims as its own. The U.S. economy grew 3.4% annualized in the third quarter and unemployment is at 4.1%, but 75% of those who voted on November 3 said the economy is doing poorly ^{[309] [165] [125]}. Electricity rose 4.4% in a year, more than overall inflation, and the 30-year mortgage rate is around 7.3% ^{[193] [307]}. Data centers entered that calculation and the campaigns understood it: they spent $137.6 million on television ads about them, 77% of them by Democrats, and of 685 campaigns taking a position in digital ads, 636 criticized data centers and 10 defended them ^[193].

Moreover, that growth generates few jobs. In the twelve months to October, the economy created 885,000 jobs and health care and social assistance contributed 519,000, almost six in ten, while information lost 119,000 positions and computing infrastructure and systems design about 26,000 each ^[315]. October was better than expected, with 165,000 jobs, 70,000 of them in professional and business services, but information and construction barely moved ^[165]. Consumption contributed 1.70 points over the last four quarters, more than AI investment ^[309], and Moody's estimates that the top 10% of earners account for about half of that spending, which it attributes to their wealth gains; the Minneapolis Fed agrees on the concentration, although it sees it as much less pronounced ^[316]. Dean Baker, of the Center for Economic and Policy Research (CEPR), an economic research center in Washington, observes that consumption grew 3.8% in the third quarter while the real wage bill rose only 0.6%, and attributes this to high-income households spending their stock market gains and to savings that are running out ^[317]. If that reading is correct, a fall in AI stocks would hit in two ways, through investment and through the spending of the wealthiest households.

Financing was already under pressure before the election. With the third-quarter reports, the combined free cash flow of Amazon, Alphabet, Meta and Microsoft was about -$26 billion, with Microsoft at +$18.6 billion, Meta at -2.1 billion, Alphabet at -9.1 billion and Amazon at -33.5 billion ^[318]. Meta had its first negative quarter and Alphabet its second in a row, and Alphabet announced that in 2027 its capital spending will increase significantly ^[318]. Meta had already placed $25 billion in debt in May ^[319]. After the reports came a wave of financing: Meta announced on October 26 about $35 billion in common and preferred stock, with nonbinding interest from Nvidia, AMD and Broadcom in buying up to 5 billion, and later placed $15 billion and €9 billion in bonds, while Microsoft placed $25 billion in bonds, Oracle $10 billion in stock and another 10 billion in bonds with coupons of up to 8.65%, and Alphabet €6 billion ^[320]. As of September 30, Amazon's long-term debt totaled $180.9 billion, versus 68.8 billion in December, and Alphabet's $125.1 billion, versus 46.5 billion ^[319]. Added to this was the cost of money, because the Federal Reserve raised its rate in September to a range of 3.75% to 4%, the first hike since 2023, and in October two governors voted to raise it again ^[321]. Goldman wrote that the biggest boost from the boom has already passed and projects 1.8% growth for the fourth quarter ^[307], and Dan Ives, an analyst at Wedbush, moved forward to 2027 the peak in big tech investment growth ^[251]. Goldman's AI basket had lost 15% in October, before any vote ^[322].

In election week, the market punished those who depend on building with debt. On Wednesday, companies that already have data centers gained, because slower construction makes their capacity more valuable and eases their cash flow, while Iren and Applied Digital, which depend on building with debt, lost ground ^[308]. On Thursday almost all of them fell, especially the suppliers, with Micron and AMD down 17% and Nvidia 9% ^{[323] [324]}. The analysts consulted by CNBC and Fox Business blamed politics, but also disappointing results, spending, rates and crowded positions ^{[251] [325]}, and Qualcomm also fell on its own forecast ^[323]. On Friday chips rebounded 4%, but in one month Nebius, IREN and CoreWeave lost about half their value and technology ended 10.7% below the S&P 500, its widest gap since 2002 ^{[326] [327]}. We believe the election was the trigger for a correction that was already under way, and no one can yet separate how much is due to each cause.

China frames growth as the diffusion of AI throughout its economy. Its “AI+” plan calls for the “intelligent economy” to be an important pole of growth by 2030 and for the country to fully enter a new stage of intelligent economy and society by 2035, and it emphasizes bringing AI to industry, agriculture and services ^[255]. The same document also calls for assessing AI's risk to employment and steering innovation toward where it creates more jobs ^[255]. China's prices also put pressure on the American bet, because if an open model charges a fraction of what a closed one does, the revenue that must pay for the investment depends on customers remaining willing to pay the difference ^[328].

We believe growth is a weak argument for defending AI's political license. The growth measured so far stays in investment, in the stock market and in the spending of the wealthiest households, and the figure of half of growth, which is gross, tells little to a voter who pays more for electricity and sees no new jobs. For investors, by contrast, this week showed that the decision of a county or a governor's office already affects the value of companies.

07

What AI already solves

ChatGPT reached 900 million weekly users in February 2026, 50 million of them paying ^[329], and Google's Gemini app passed one billion monthly users in October, according to Alphabet ^[330]. More than 70% of ChatGPT's messages have nothing to do with work, and nearly 80% of conversations are practical guidance, information seeking and help with writing, according to a study by economists at OpenAI and Harvard University ^[331]. On Anthropic's Claude, personal use rose from 35% to 42% of conversations between November 2025 and February 2026 ^[332]. Microsoft estimates that in the second quarter of 2026, 18.8% of the world's population aged 15 to 64 used generative AI, 28.8% in the Global North and 16.2% in the Global South ^[333]. In the United States, 66% of small businesses use some form of AI, eight points more than in 2025, according to the U.S. Chamber of Commerce ^[334]. In the second quarter of 2026, 39.2% of employed adults aged 18 to 64 used it for work in the previous week, and the time they report saving equals 2.2% of all hours worked, up from 1.6% in 2024, according to the survey published by the Federal Reserve Bank of St. Louis ^[335].

In software, the change is already massive. Between May and July 2026, 90% of the more than 15,000 professional developers JetBrains surveyed worldwide used AI agents at work at least weekly, and 68% every day. Anthropic's Claude Code was the most used tool, by 39% of developers (47% in the United States), up from 18% in January; OpenAI's Codex rose from 3% to 16%, GitHub Copilot fell to 21% and Cursor to 12% ^[336]. On average, those developers say agents write about 47% of their code in full, and one in five no longer writes a single line without AI help ^[337]. The research firm SemiAnalysis estimated that in February Claude Code authored 4% of public commits on GitHub ^[338], and Anthropic reported that month that the tool's run-rate revenue had passed $2.5 billion, more than double its level on January 1 ^[339]. Codex passed 5 million weekly users in June, and about 20% of them are not developers but analysts, marketers, designers, researchers, investors and bankers ^[340].

People who cannot code are already building applications. On Lovable, which builds an application from a description written in plain language, 80% of builders hold non-technical roles, more than half of those surveyed say they are building a business, and what they publish receives an average of 720 million visits a month, according to the company's own data ^[341]. GitHub now has more than 225 million developers, and in September 2026 more than 40% of pull requests started with a first draft written by an agent, up from 17% in March ^[342].

Models now pass tests that until recently only the best could solve. In July 2025, a version of Google DeepMind's Gemini Deep Think solved five of the six problems at the International Mathematical Olympiad and scored 35 of 42 points, a gold medal graded by the olympiad's coordinators with the same criteria applied to students ^[343]. In September, at the World Finals of the International Collegiate Programming Contest in Baku, the same model solved 10 of 12 problems, a gold-medal performance that only four of 139 teams reached, including one problem no university team solved ^[344]. AlphaEvolve, a Google DeepMind agent, found the first improvement in 56 years for multiplying 4×4 complex matrices and a rule for scheduling jobs on Google's servers that recovers, on average, 0.7% of all its compute ^[345].

Science is advancing across many disciplines, and in some the scale already exceeds what any person could do without AI; protein biology and mathematics are two examples. AlphaFold predicted the structure of virtually all 200 million known proteins, work that received the 2024 Nobel Prize in Chemistry ^[346], when half a century of crystallography and cryo-electron microscopy had gathered some 262,000 experimental structures, at a pace of about 17,600 a year ^[347]. In mathematics the change is one of scale: on October 6, 2026, OpenAI published 722 manuscripts with new results, obtained by posing some 4,000 open problems to an internal model, and about 42% have a computer-verified proof in the Lean language ^[348]. From the catalog of problems left by the mathematician Paul Erdős alone, that release addresses 25, in 1,347 pages ^[349], after an AI system solved the first of them on its own in January ^[350] and another model refuted a conjecture Erdős posed in 1946 in May ^[351]. In September, some 10,000 OpenAI agents produced in 88 hours a proof that the Navier-Stokes equations, which describe the motion of fluids, can develop a singularity when an external force acts on the fluid ^[352]. On October 26, OpenAI also announced that its models proved the local smoothing conjecture, one of the central propositions of harmonic analysis, in every dimension, when it had only been proved in three, and that they refuted the Exponential Time Hypothesis, on which much of the theory of computation has been built since the late 1990s ^[353]. The caveats weigh as much as the figures: almost none of this has been peer reviewed, the editor of the Erdős catalog describes the texts as poorly written, OpenAI withdrew three manuscripts the next day over a sign error ^{[348] [349]}, many of the problems had remained open because no one had attended to them rather than because of their difficulty ^[354], and the Navier-Stokes result is not the version of the Clay Institute prize, is still under review, and contests priority with two mathematicians who reached the same result, also with the help of AI ^[352]; Ryan Williams, a specialist at MIT, said his team is still checking whether the refutation of the hypothesis holds up ^[353]. AI expands what can be attempted, and verifying what it produces still depends on specialists able to review it.

In Sweden, a trial with 105,934 women showed that reading mammograms with AI support detects 29% more cancers with no increase in false positives ^[355], cuts radiologists' workload by 44% ^[356] and raises sensitivity from 73.8% to 80.5% without increasing the cancers that appear between one screening and the next ^[357]. Google DeepMind's GenCast outperforms the ensemble forecast of the European Centre for Medium-Range Weather Forecasts, the world's best at medium range, on 97.2% of 1,320 measures, and produces a 15-day forecast in eight minutes on a single processor, when the traditional method takes hours on a supercomputer ^[358]. A Google model forecasts extreme floods on rivers without gauging stations five days ahead, with the reliability that same-day warnings used to have, against a risk of severe flooding that affects about 1.5 billion people ^[359].

08

The race with China

The United States and China are competing over who will lead the transition to what the World Economic Forum named the fourth industrial revolution in 2016 ^[145], and artificial intelligence sits at the center of that transition because it cuts across every other industry. In July 2025, the U.S. government published a plan titled “Winning the Race,” which presents global technological dominance as a national security imperative ^[360]. China responded a month later with its “AI+” opinion, which aims to bring intelligent terminals and intelligent agents, AI programs that carry out tasks on their own, to more than 70% penetration by 2027 and more than 90% by 2030 ^[361].

At bottom, both are seeking an advantage the rival can no longer close. The thesis that dominates in American labs holds that once AI can do AI research, each generation of models will accelerate the next, so that whoever reaches that point first, known as recursive self-improvement, could compress years of progress into months. Leopold Aschenbrenner, a former OpenAI researcher, wrote in 2024 that hundreds of millions of systems capable of automating AI research would give a decisive economic and military advantage ^[362]. Dario Amodei, CEO of Anthropic, describes “a country of geniuses in a datacenter” and proposes that a coalition of democracies use that advantage, even if temporary, to secure robust military superiority ^[139]. OpenAI even set itself dates: Sam Altman announced in October 2025 an automated AI research intern for September 2026 and an automated AI researcher for March 2028, and on September 29 he said the first already exists ^{[363] [364]}. If the thesis is correct, whoever arrives second will fall behind in weapons, intelligence, cybersecurity and trade, and the gap will widen on its own. That is why Dan Hendrycks, Eric Schmidt and Alexandr Wang draw a troubling conclusion: a state that sees its rival on the verge of getting there would have reasons to sabotage it first, with cyberattacks or even attacks on its data centers, in a logic of deterrence they compare to the nuclear one ^[365].

The thesis already appears in Anthropic's filings with the SEC. In the November 2 amendment to its prospectus, the company warns that, based on its internal estimates and its expectations for future models, it may soon reach the threshold that its Responsible Scaling Policy, the internal rules that set which safeguards each capability level requires, calls “dramatically accelerated progress attributable to AI,” defined as advancing in one year what would normally take two; it clarifies, however, that none of its current models crosses it and says it is studying requiring human approval for certain research decisions ^{[366] [367]}. By its own measurements, in August Claude was directing 26% of its AI research and development work, up from less than 1% in February, and some 30,000 agents were simultaneously carrying out research and engineering tasks on its main internal platform ^[368]. METR, an independent organization that evaluates the risks of AI models, audited it for eleven weeks and estimates that its pace of progress reached about double what it had been between March 2024 and October 2025, with a range of 1.3 to 3 times; it attributes perhaps 80% of that acceleration to AI and rules out that it comes from more training compute, that is, from more processing capacity to train the models ^[367]. METR also warns that Anthropic redacted most of the evidence METR considered most informative and that it withdrew, midway through the work, the verification of how it allocates its compute. It also judges overstated the claim in Anthropic's public risk report that the threshold had not been crossed, because it finds it plausible that it was crossed before July 15, although most of the researchers consulted lean toward thinking it was not ^[367]. Sam McCandlish, Anthropic cofounder and head of its scaling policy, replied that disclosing what drives the acceleration would amount to giving others a guide to go faster, and announced for the coming weeks a report of its own on where the company stands relative to that threshold; Daniel Kokotajlo, of the AI Futures Project, a group that publishes forecasts on AI progress, wrote instead that, if Anthropic is about to reach it, timelines to superintelligence, an AI that surpasses people at almost every intellectual task, could be very short ^[369]. An internal model that finished training between September and October exceeds 50% on every test Anthropic uses to measure progress, and METR fears it will not be able to keep measuring it ^[367].

In the November 10 amendment, Anthropic says it determined that it has met the threshold, based on what it expects from future models and its overall rate of progress rather than on what it has observed from current models, and acknowledges that its ability to measure that acceleration is limited and that in the past it did not identify a significant acceleration until after it had occurred ^[97]. The same day, the AI Futures Project moved up its forecasts: all its authors now expect, absent a substantial deliberate slowdown, a superhuman AI researcher in 2027, an AI that matches the research taste of the best researchers at a lab and can fully automate its coding. Kokotajlo, the one who expected the shortest timelines, put his median at December 2028 in August and now puts it at May 2027; three months earlier, his most aggressive model gave less than a 4% chance that by November progress would be sped up at least 1.7 times by AI, the figure the group now derives from the METR report. The group concludes that it is urgent to pause or substantially slow frontier development, ideally with an industry-wide pause followed by an international agreement, because companies are not on track to align such a researcher with human values and goals ^[98].

The labs, for their part, are already holding back that capability in their products. On November 9, Anthropic launched Fable 5.2 for the general public and Mythos 5.2 only for verified organizations, and explains that in training it deliberately limited certain AI research capabilities, such as optimizing accelerator code and designing model architectures, to about the level of Opus 5.5 ^[370]. OpenAI reported on November 4 that GPT-6 Astra met about 40% fewer of the criteria of the same programming task when it was presented as safety research than when it was presented as capabilities work, without its monitors, the automated systems that watch the model's behavior, flagging it and without the company finding a conclusive cause ^[371]. We believe that if the models that will do the research perform worse when that research is about safety, oversight of self-improvement faces a problem that no threshold solves. Since October 15, six METR employees have been working inside Anthropic as independent evaluators, with access to its systems; Anthropic cannot veto what they publish, but it can require that categories defined in advance be redacted, and McCandlish clarified that the November 5 evaluation predates and is unrelated to that arrangement ^{[372] [369]}.

The same thesis the labs invoke to seek support is fueling a bill in the U.S. Congress to ban superintelligence. On September 23, Bernie Sanders introduced in the Senate the Artificial Superintelligence Prohibition Act (S. 5493), which he announced together with Greg Casar, chair of the Congressional Progressive Caucus ^{[373] [374]}. The bill permanently bans developing or deploying systems that exceed human cognitive capacity or that can plan and carry out the destruction or subjugation of humanity, pauses the most advanced systems until a new Department of AI sets safety rules, requires that department's approval before deploying them and punishes noncompliance with dissolution of the company and up to 20 years in prison ^[374]. It also takes up the statement published in October 2025 by the Future of Life Institute, a nonprofit dedicated to reducing extreme risks from advanced technologies, which calls for prohibiting superintelligence until there is scientific consensus on its safety and broad public support, and which has more than 78,000 signatories, among them Geoffrey Hinton, Yoshua Bengio, Steve Wozniak and Richard Branson ^[375]. TechRepublic warns that such a broad definition would reach systems far less autonomous than those the bill seeks to prevent ^[376]. The bill remains in the Commerce Committee and, with a Republican majority, had little chance of advancing. We believe it will carry more weight from January, because with the Democratic majority voters elected on November 3 it may reach hearings, and every lab that talks about recursive self-improvement will hand it arguments.

The labs sell to three customers at once: the consumer, with ChatGPT, Claude or Gemini; large companies, which pay to bring the models into their software and processes; and the state. In July 2025, the Pentagon's Chief Digital and Artificial Intelligence Office awarded Anthropic, Google, OpenAI and xAI contracts of up to $200 million each to build agents for warfighting, intelligence and administrative tasks ^{[291] [292]}. Since June 2025, Anthropic has offered Claude Gov models built only for national security customers, which, according to the company, are already used in classified environments by agencies at the highest level of U.S. national security, without naming them, while OpenAI gathered its public offering under OpenAI for Government, with models tailored to national security and deployments at Los Alamos, Lawrence Livermore and Sandia ^{[377] [292]}. On October 27, Anthropic gave federal, state and local governments a 40% discount on its products, with no minimum spend, even though the Department of War keeps it designated as a supply chain risk ^[378]. That third customer gives the lab an argument few software companies have, because as a supplier to national defense it can ask for energy, chips, permits and special regulatory treatment. At the same time it exposes it to what the state demands in return, which is the discussion in the section “What the state can demand of a supplier.”

Arvind Narayanan and Sayash Kapoor, of Princeton, reject the leap thesis: they argue that AI is a normal technology, whose effect depends on people, companies and institutions changing the way they work, and that this diffusion takes decades, as electricity's did; superintelligence, as it is usually understood, strikes them as an incoherent concept ^[379]. China does not bet everything on that leap either, since its “AI+” opinion measures success by AI's penetration of the economy, and Will Rinehart, of the American Enterprise Institute, sums up China's strategy as a race to embed AI in the economy as fast as possible ^[380]. Other analysts, such as Matthew Johnson, of the Jamestown Foundation, a Washington think tank specializing in China and Eurasia, argue that China is pursuing both at once, and in fact its 15th Five-Year Plan, presented in March, calls for exploring paths toward artificial general intelligence, an AI capable of matching people at most intellectual tasks ^{[381] [382]}. We believe the American industry rests its political argument on the leap thesis, while the diffusion thesis better describes what voters see today. If the leap comes, whoever arrives first wins; if it is slow or never comes, whoever has diffused the technology through its economy wins, which is the historical advantage described by Jeffrey Ding, a political scientist at George Washington University ^[383]. In either case, the United States needs its own society to let it keep building.

The two countries made different bets. The United States bet on the most capable model, closed and financed with private capital, with $285.9 billion in 2025, 23 times China's $12.4 billion, according to Stanford's AI Index ^{[384] [385]}. The report itself warns that this comparison understates China, because state funds do not appear in private capital databases ^[385]. The National Venture Capital Fund, launched in December 2025, starts from 100 billion yuan (about $14 billion) in special Treasury bonds to mobilize 1 trillion yuan (about $142 billion) over twenty years, with AI as its first target, and in 2026 Beijing and Guangzhou created municipal funds of 100 billion and 20 billion yuan (about $15 billion and $3 billion) ^{[386] [387] [388]}. China, by contrast, bet on open, low-priced, good-enough models, and on putting them into industry before anyone else. An open model publishes its weights, the numerical values that define it, also called parameters, so that anyone can download it, modify it and run it on their own hardware. Per million output tokens, the units of text by which a model's output is measured and billed, Anthropic's most expensive model charges $50, Moonshot's Kimi K3 $15, Z.ai's GLM-5.2 $4.40 ^[328] and DeepSeek V4 Pro $3.96 at peak hours and $1.98 off-peak ^[389]. In September DeepSeek released V4.1 Flash under the MIT license, which allows it to be used and modified freely; according to the company it outperforms V4 Pro, and it charges $1.20 per million output tokens at peak hours and $0.60 off-peak ^[389]. The United States also has open models, although they lag behind. Reflection, which raised $2 billion with Nvidia among its investors, released Beam in October, with 501 billion parameters, under the Apache 2.0 license, also open; in the company's own table, Beam scores 80.1 on Terminal-Bench 2.1, a benchmark of command-line programming tasks, while the top four spots are Chinese, with DeepSeek V4.1 Flash at the top, at 90.6 ^[390]. Chip export controls forced those labs to squeeze their compute; they also pay less for electricity thanks to state investment in generation and accept thin margins to gain market share ^[328]. The bet is already reaching American companies: in one week in July, between 57% and 63% of the tokens they consumed on OpenRouter, a platform that provides access to hundreds of models from different providers, depending on the measurement, came from Chinese models, a volume figure on a single platform; Coinbase cut its AI spending in half by switching to Kimi and GLM, and the coding model of Cursor, the AI code editor, is built on Kimi ^{[328] [391]}.

That bet rests on a broader base than the conversation in the United States acknowledges. China authors 23.2% of the academic papers on AI published worldwide and receives 20.6% of citations, compared with 12.6% for the United States, and accounts for 69.7% of patent applications ^[385]. Of the hundred most cited AI academic papers of 2024, 41 are Chinese, up from 33 in 2021, although the United States retains the highest-impact patents and more notable models, 59 to 35 in 2025 ^[392]. Nvidia CEO Jensen Huang summed it up in December 2025: “50 percent of the world's AI researchers are Chinese,” 70% of the previous year's AI patents were published by China, nine of the ten top science and technology schools are in China, and “we used to lead most of them; now they lead most of them” ^[393]. The figure on institutions matches the Nature Index, which measures each institution's output in the leading scientific journals: in its 2026 table, nine of the top ten are Chinese, with the Chinese Academy of Sciences and Zhejiang University in front and Harvard in third place ^[394]. The flow of AI researchers to the United States has fallen 89% since 2017, and China installed 295,000 industrial robots, almost nine times the United States' 34,200 ^[385]. In 2025 it also added 542 gigawatts of electric capacity ^[252]. In March, the gap between the best American model and the best Chinese model was 2.7% ^[385].

The American industry and its government respond that much of that closeness is explained by unauthorized copying. The technique is called distillation and consists of training a model on the outputs of another, which is legitimate when the model's owner authorizes it. In February, Anthropic accused DeepSeek, Moonshot and MiniMax of doing so with Claude without its consent, in violation of its terms of use and its access restrictions; since Anthropic does not sell Claude in China, the three labs reached it through resellers, with some 24,000 fake accounts and more than 16 million exchanges ^[395]. On September 8, the NSA, CISA, the federal cybersecurity and infrastructure agency, and the FBI accused six Chinese companies, among them Alibaba and Z.ai, of extracting without authorization billions of tokens from Claude, GPT, Gemini and Grok since late 2024, probably with the knowledge of the Chinese government, and called distillation the core of their strategy ^[396]. China's Ministry of Commerce replied the next day that the accusations are baseless, that distillation is a neutral technique that American companies also use and that the United States seeks a monopoly over the industry ^[397]. In October, Anthropic added that Zhipu, the company behind Z.ai, distilled cyberattack capabilities for GLM-5.3 and that its new blocking reduced by between 90% and 99% what distillers manage to extract ^[398]. DeepSeek V4's own technical report admits it is three to six months behind the frontier, that is, the most advanced models, while the Council on Foreign Relations estimates the American lead at about seven months and CAISI, the Center for AI Standards and Innovation of the Department of Commerce, in its section on China, at about four ^[399]. We believe unauthorized copying lowers the training cost for whoever is behind without putting it ahead, and that the debate over it distracts from the front on which China is in fact winning, which is price and adoption. Michael Horowitz, of the same Council, argues that the decisive competition is the diffusion of open, good-enough, low-priced models, above all in the Global South ^[399]. He thus agrees with Jeffrey Ding's historical thesis, according to which the powers that won the industrial revolutions were the ones that diffused the technology through their economies, even if they had not invented it ^[383].

In cybersecurity, the race has already become concrete. October 2026 is being called Black October because of the accumulation of serious vulnerabilities disclosed within a few days, several of them found with the help of AI ^[400]. The most serious, PHP2Shell, was found in August by Ryan Austin, a researcher at Calif, a California cybersecurity firm, using an OpenAI model, and the cybersecurity firm Wiz found related flaws with OpenAI and Google models. It affected the language that runs about 69% of websites, its details leaked before the October 21 patch, and as a result the Wikimedia Foundation, the Internet Archive and government sites in Europe and Canada went offline to protect themselves ^[401]. That same day, Microsoft attributed to Storm-0940, a group aligned with the Chinese government, a campaign that, with credentials stolen from ConnectWise Cloud Backup, tried to break into more than 2,250 Microsoft 365 accounts in government, defense, aerospace, finance, energy and semiconductors by means of ARTEX, an open tool that coordinates AI agents to carry out an intrusion from start to finish; in several cases it went from the first attempt to privileged access, that of an administrator, in under three hours, and the Chinese foreign ministry denied any knowledge of the matter ^[402]. The millions of attempts produced few successful attacks, in part because companies patched in time, and Axios compared it to the year 2000 ^[403]. The labs already restrict access to their most capable models in this area. Google gave Gemini 4 Carbon on October 28 only to a group of cyber defenders, through its Fairwind program, is taking part with it in the U.S. government's voluntary process for testing models before release and has not set a date for developers and consumers ^[404]. OpenAI, for its part, rated as “Critical” in cybersecurity, the highest level on its risk scale, GPT-6 Astra, which it launched in September, and GPT-6 Cyber, from November 5, which it reserves for selected customers ^[405]. Nine days after the patch, attackers used an internal Binance AI tool to reach the systems it uses to sign its transactions and made off with about $6.34 billion, a theft the FBI attributes to North Korea ^{[406] [407]}. We believe the episode gives arguments to both sides, because the labs can show that their models find vulnerabilities before attackers do, while their critics can point out that the same capability, distilled or open, is already in the hands of groups aligned with a rival government.

A shortage of chips and revenue is holding China back. DeepSeek cannot serve V4 Pro to most of its customers for lack of chips ^[399]. Its labs together bill about a tenth of what OpenAI and Anthropic bill, and openness erodes their pricing power: when Z.ai released the weights of GLM 5.3 Flash, its share of that model's tokens on OpenRouter fell from 88% to 22% in twenty days ^[408]. The STAR 50 index, which groups its chip and AI champions, has fallen more than 40% since June; Chinese stocks have lost about $3.5 trillion since May, according to Bloomberg, and the Chinese government injected $19 billion into exchange-traded funds in one week ^{[409] [410]}. Analysts debate how much of that drop is due to an AI bubble and how much to weak consumption, housing and taxes, but chip demand continues, and CXMT, China's main memory chip maker, earned 54 billion yuan (about $8 billion) in the third quarter while its stock fell ^{[409] [411] [388]}. Outside China, moreover, few trust its government to regulate the technology: in Pew's 25-country survey, the median is 27%, compared with 37% for the United States and 53% for the European Union ^[412]. Wall Street also has doubts about the American side, as the Nasdaq hit a record on October 6 and ended the month near its three-month low, with investors asking when the spending on data centers will pay off ^[413].

In perception, by contrast, the comparison favors China. In Ipsos's 2026 survey of 32 countries, China is the most excited about AI (83%) and among the least nervous (40%), while in the United States only 33% say it excites them and 64% that it makes them nervous, among the four least excited and the four most nervous ^[412]. In China, more than 80% of employees use AI at work regularly, compared with 58% worldwide, and 78% expect it to save them time; in the United States only 33% believe it will improve their job, and in Pew's 2025 global survey just 31% trusted their own government to regulate it, the lowest figure of the 25 countries ^[412]. China imposes its diffusion strategy from the state and finds a population willing to adopt it, while in the United States every data center, every layoff attributed to AI and every lawsuit is debated in public and voted on. An authoritarian government is more efficient at developing the technology, because it decides where a data center goes and how much energy it receives without a vote, a court or a strike stopping it. We believe this is the disadvantage the American industry least acknowledges.

The U.S. economy already depends on that buildout. According to our calculation with the national accounts of the Bureau of Economic Analysis (BEA), investment in computing, software and data centers contributed 1.03 of the 2.08 points of growth over the last four quarters, although this is a gross figure, because imports of that same equipment subtract another portion ^[309]. Yet that technology worries many of those who live off that growth, and the hegemony argument says little to a citizen who does not feel that race as their own.

In the United States, Democrats won the House of Representatives and the Senate on November 3, as well as the governorships of Texas, Georgia, Pennsylvania and Ohio ^{[414] [200]}. The economy decided more votes than AI: in the NBC exit poll, 39% named it their top issue, 75% said it is doing badly and 65% that President Trump's policies made it worse ^[125]. Before voting, 84% of registered voters considered the economy very important to their vote, while only 44% said the same of AI, which ranked eleventh of sixteen issues ^[415]. Trump's approval stood at 30% in Pew and reached 38% in CBS exit polls ^{[416] [414]}.

AI entered that vote through jobs and data centers. In the NBC exit poll, 76% oppose a data center in their community ^[125]. In the CNBC survey, 56% said the issue weighed crucially or importantly in their decision, and 65% of that group backed a Democrat ^[213]. In 41 local votes in ten states, voters backed restrictions in all but one ^[21]. Sherrod Brown won back an Ohio Senate seat after calling his opponent the face of data centers ^[417], and Gina Hinojosa defeated Greg Abbott in Texas, the state with 27% of the country's planned gigawatts, after promising to halt all approvals until the state legislature acted ^{[201] [418]}.

The stock market celebrated the result one day and punished it the next. On Wednesday the 4th, the S&P 500 rose 1.6% and Amazon and Apple soared, but on Thursday the 5th the index fell 4.36%, the Nasdaq 5.52% and semiconductors about 12%, their worst drop since 2020 ^{[323] [308] [419] [324]}. On Friday they partly rebounded, and the week closed with the S&P 500 down 2.7% and the Nasdaq down 3.4% ^[326].

In the United States, resistance to AI is as great as the enthusiasm it arouses. In politics, what people believe is happening to them weighs as much as what the data series record, and in that perception China has the advantage today.

09

Teaching AI

Whoever teaches people to use AI largely decides how teachers, doctors, lawyers and technicians see it, and both governments have turned that into policy. In the United States, Executive Order 14277, of April 2025, set AI literacy as policy, created a task force at the White House, directed that AI be prioritized in federal grants to train teachers and in apprenticeship programs, and asked agencies to partner with companies in the sector to produce school materials; in June, more than 60 organizations signed the pledge that accompanies the order ^[420]. In China, the Ministry of Education ordered at the end of 2024 that AI instruction be strengthened in primary and secondary schools and that teacher training be included in its plans, and Beijing has required at least eight hours per school year since the fall of 2025 ^[421]. The difference lies in who writes the content, which in China is the State, while in the United States the government largely entrusts it to the companies that sell the tool.

The pharmaceutical and medical device industries show that training professionals changes what they use. A review of 36 studies by Mitchell and colleagues finds an association between industry payments and prescribing; a payment from Intuitive Surgical, the maker of the surgical robot, is associated with 9.9 more points of use per surgeon, and that manufacturer reports 76,000 surgeons trained worldwide ^[422]. In AI there are already similar programs, such as the national academy of the American Federation of Teachers, which, with $23 million from Microsoft, OpenAI and Anthropic, is aimed at 400,000 educators and gave its first course in March ^[423], Anthropic's programs for teachers with the same union, and for-credit courses with professional associations ^[424]. None of these programs publishes how much the income or employment of those who took them improved. Outside the industry the criticism is direct: education technology analyst Audrey Watters wrote that teaching teachers Microsoft's tools helps Microsoft more than students, and Benjamin Riley, founder of Cognitive Resonance, a think tank he created in 2024 to study AI, warned that these partnerships seek market share and can devalue the teacher's work ^[425].

Training, however, is not the same as winning defenders. The studies of the medical industry measure payments and prescribing, and none measures political advocacy, and speaker programs, the channel that moved prescribing the most, ended in an Inspector General alert and a $678 million Novartis settlement ^[426]. The authors of the robot study also warn that these relationships can compromise the value of care ^[422]. The closest recent experiment points the other way: Yotam Margalit and Shir Raviv randomly assigned more than 1,500 workers a human boss or an AI boss, and the experience, good or bad, changed how they worked without changing what they thought about using AI in public policy; information, by contrast, did move those opinions, and information about risks lowered support ^[427]. Training a translator in the tool that took away his or her assignments may not generate loyalty, and we found no evidence of successful retraining in analogous programs.

Preferential access has precedents and counterexamples, because there are better plans for U.S. university students than for those in other countries and, at the same time, more generous free offers outside the U.S. ^[428]. The token price for the federal government drew a protest from Ask Sage, a competitor that, among other reasons, argued it would lock in the customer ^[429]. A program whose declared goal is to train defenders is exposed on two fronts, since spending aimed at influencing the public on legislative matters is not tax deductible (26 U.S.C. §162(e)), and the partners, unions and associations, withdraw if they read that they are considered a political base.

China shows that adoption does not need to win over professionals one by one. In a University of Melbourne survey across 47 countries, more than 80% of employees in China, India, Nigeria, the UAE and Saudi Arabia said they use AI regularly at work, compared with about half in North America and Europe ^[430]. There it is pushed by the State, with its “AI+” program ^[255], and by employers.

Training people on the tool can also take away the skill it replaces. David Krakauer, president of the Santa Fe Institute, distinguishes tools that leave competence with the user, such as the abacus, from tools that replace it, such as GPS, which leave the user worse off than before once they are taken away, and he warns that the latter lead to complete dependence on whoever provides the capability ^{[431] [432]}. There is evidence from medicine: at four centers in Poland, 19 experienced endoscopists saw their adenoma detection rate in colonoscopies without AI fall from 28.4% to 22.4% after working with it, although the authors caution that the sample is small ^[433].

We believe the political effect of a training program would be small, judging by what the experiment shows. A program written by whoever sells the tool also runs the risk its critics point out, that teachers read it as advertising. In our judgment, adoption depends more on who hires than on who teaches.

10

Depending on a lab

The dependence between the labs and those who build on them is already producing cases. OpenAI will stop serving Cursor on November 12, with 76 days' notice, because SpaceX bought the company and OpenAI says it cannot trust it to respect its terms ^[434]. In January, Anthropic had cut off xAI, which used its models through Cursor, and it also blocked the use of its subscriptions in third-party tools ^[435]. Anthropic's own prospectus acknowledges that it competes with some of its customers, especially in software as a service ^[366]. Brookings proposes a rule prohibiting different treatment of similarly situated customers ^[435]. Today there are published model retirement notices and public prices, but no advance notice for cutting off a customer ^[436]. In similar businesses there are models that can serve as a reference: the European Union requires platforms to give 30 days' notice and reasons before cutting off a business user, although a model API may fall outside that, and sets a maximum period of 30 days to switch cloud providers; China, for its part, prohibits dominant companies from abusing their position through data, algorithms and platform rules ^[437]. Amazon sued Perplexity to prevent the shopping agent in its Comet browser from accessing its customers' accounts, and in March obtained a preliminary injunction. In August, however, the Court of Appeals for the Ninth Circuit vacated it, because it considered Amazon unlikely to prevail and treated the agent as a tool of the user, although it limited the ruling to the current record, and the case remains in the trial court ^[438].

The cloud showed that a platform can profit from its customers' success. AWS gave $7 billion in credits to more than 280,000 startups, and the cloud is today the largest part of Amazon's profit, with margins of 36% to 39% ^[439]. Platforms report $4 to $7 of activity for every dollar of their own, although in studies they commissioned themselves ^[440]. However, there is no measurement of how many supported startups survive, AWS won mainly with the migration of large enterprises, and a cloud lends on a high margin and with spare capacity ^{[441] [439]}. A lab, by contrast, buys compute 12 to 18 months in advance, so guaranteeing capacity to a startup competes with the paying customer ^[442]. The labs' startup programs are small, and the largest fund tied to one equals about 5.7% of the credits of the associated cloud ^[443].

The labs' accounts do not settle whether it is better to be a platform or an application either. Anthropic, whose public offering is underway, had operating income of $974 million in the second quarter and returned to a loss in the third, at -$567 million on revenue of $16.956 billion, with compute at 58% of revenue; in nine months it took in $33.242 billion, lost $1.26 billion on operations and committed to buy about $895 billion of compute, most of it before 2031 ^[366]. OpenAI burned about $3.7 billion in the first quarter of 2026, according to the press, and its enterprise revenue reportedly surpassed its consumer revenue ^[444]. SpaceX's AI segment, which includes xAI and the X network, went from a positive result in 2023 to a negative one in 2025, in step with compute ^[445]. It follows from these figures that scale and compute purchases weigh more on results than the business mix.

In 2025, the U.S. Federal Trade Commission (FTC) flagged the practice of tying support to exclusivity or to spending on a partner's cloud, which antitrust law can reach, and below-cost pricing at large scale also raises the question of predatory pricing ^[446]. The current administration ordered a review of FTC investigations so that they do not unduly burden AI innovation ^[447], although in October the FTC itself opened a civil investigation into Anthropic on consumer protection, data, privacy and security ^[366]. With a Democratic Congress, moreover, the labs' competition with their own customers becomes a subject for hearings.

China has already punished both practices. In 2021 it fined Alibaba 18.228 billion yuan, 4% of its domestic sales, for requiring merchants not to sell on rival platforms ^[448]. In September 2025, the State Administration for Market Regulation (SAMR), its competition authority, preliminarily concluded that Nvidia had violated the conditions under which its acquisition of Mellanox was approved, among them not tying the sale of its chips to that of its networking equipment, and it did so three days after the United States added 23 Chinese companies, most of them semiconductor firms, to its entity list of restricted exports ^[449]. China thus uses the same instrument to order its market and to respond to a U.S. supplier.

The dependence also rises a level, from the labs to the clouds. FTC staff concluded in January 2025 that the partnerships of Microsoft with OpenAI and of Amazon and Google with Anthropic can lock the labs into a cloud and give the cloud sensitive information ^[450]. Amazon is at once Anthropic's investor, creditor and supplier: it invested $8 billion in convertible notes and $10 billion in preferred stock, opened a facility for it that becomes available as AWS delivers compute, and in the second quarter the two expanded their commercial commitment by more than $100 billion over ten years. As of September 30, those positions totaled $238.7 billion on Amazon's balance sheet, at estimated values without a market price, and their upward adjustments contributed 43% of its pretax income over nine months ^[451]. Amazon also funds the rival, since in 2026 it invested $50 billion in OpenAI preferred stock, and OpenAI's compute contract with AWS grew by $100 billion over eight years; Microsoft, for its part, billed OpenAI about $10 billion in the third quarter, around 11% of its revenue ^[319]. On top of this, the U.S. platform already has a substitute, because Chinese models are the lower-priced alternative to which U.S. companies are migrating ^[328], and Qwen, Alibaba's family of open models, was the base of 69% of new open-model fine-tunes in February, that is, of the versions others adapt from a published model ^[452]. OpenAI released open models again in 2025 for the first time in almost six years, a decision that Sequoia and Stanford's analysis attribute to that competition ^[452].

We believe, with our interest in plain view, that this dependence is mutual and poorly governed, and that on this front the cost of a verifiable rule is lower than that of the harm. A customer who is cut off already has somewhere to go, and each cutoff without notice pushes another toward a free Chinese model. Since Psiquepol + AI builds on models from several providers, this conclusion favors us.

11

The bubble risk

The American public does not yet see a bubble. In December 2025, 74% expected the boom to continue and 15% anticipated a crash, and no major pollster asked again in 2026, although the University of Michigan consumer survey recorded on November 6, in open-ended responses, “growing anxiety over a possible AI bubble” ^[453]. Those who watch over financial stability, by contrast, have been naming it for a year. The Bank of England, the IMF and the Federal Reserve have included it in their reports since October 2025, and in May 2026 half of those surveyed by the Fed identified it as their main concern ^[454].

In November the risk moved to the credit markets. The cost of insuring against a default by CoreWeave, which rents compute to the labs, reached a record 1,289 basis points on November 5, and its 2031 and 2032 bonds yield 15% to 16% ^[455]. Oracle is one notch above speculative grade, and S&P projects its free cash flow to be negative by about $42 billion in its next fiscal year ^[456]. The four big tech companies had negative combined free cash flow in the third quarter, with Amazon at -$42 billion over twelve months, and only Microsoft trimmed its plan, lowering its fiscal year 2027 capital spending from $175 billion to $170 billion because of delays at its outside compute suppliers and not for lack of demand ^{[318] [457]}.

Anthropic's initial public offering is the most visible test. In the November 10 amendment, the company set a range of $70 to $78 per share for 825 million shares, reported run-rate revenue of about $81 billion as of October 31, up $10 billion from a month earlier, and added the commitment of AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC to buy $26 billion in shares at the offering price ^[97]. With the shares it will have outstanding when the offering closes, excluding options, the midpoint of the range values it at about $970 billion, by our own calculation, below the $1.5 trillion that, according to Reuters, investors saw as a ceiling days earlier and the $2 trillion once mentioned ^{[97] [458]}. The third-quarter loss and the post-election market drop thus arrive just before it goes out to find buyers. Its rival did not get what it sought either, as OpenAI closed about $11 billion on November 6 at a valuation of $1.25 trillion, according to Bloomberg ^[459].

Opposition to a bailout crosses party lines, although only Democrats have taken it to a resolution. After a remark by Sarah Friar, OpenAI's chief financial officer, about federal guarantees, David Sacks, then the White House AI czar, said there would be no bailout, and Elizabeth Warren demanded explanations ^[460]. On October 29, Greg Casar, with Maxine Waters and 11 other Democrats, introduced in the House a nonbinding resolution that rejects any bailout of insolvent AI companies by the Fed, the Treasury or the FDIC, the federal agency that insures bank deposits, and calls for them to go through ordinary bankruptcy; it is backed by the AFT, the teachers' union, Americans for Financial Reform and the Future of Life Institute ^[461]. Among Republicans, Ron DeSantis has publicly opposed it since 2025, but lawmakers have shown no interest in the issue ^[462]. The industry publicly rejects federal guarantees while asking for federal land and permits, and that gap will be debated ^[463]. A week before its chief financial officer's remark, Chris Lehane, of OpenAI, had written to Michael Kratsios, director of the White House Office of Science and Technology Policy, to ask that the 35% tax credit be extended to servers, data centers and the grid, and suggested loans or guarantees ^[464]. The Federal Reserve can lend in an emergency only through programs with broad eligibility, so a program created to save a single company from bankruptcy is excluded. In the opposite direction, Executive Order 14318 directs support for data center projects through loans, guarantees and offtake contracts, and the Defense Production Act allows guarantees to be granted when credit is not available on reasonable terms ^[465]. That is why we believe the practical debate will be over which advance guarantees are acceptable. Meanwhile, measured revenue is growing fast: Anthropic went from annualized revenue of $47 billion in May to about $81 billion on October 31, and Nvidia sells $96.2 billion per quarter with a gross margin of 75% ^{[466] [97]}.

Critics also point out that part of the demand is circular. The letter of intent under which Nvidia would invest up to $100 billion in OpenAI, which in turn buys its chips, became a $30 billion investment in the March round, and Nvidia guarantees up to $105 billion of data center leases for an OpenAI subsidiary in Ohio ^[467]. AMD gave OpenAI rights to 160 million of its shares, about 10% of the company, and Nvidia guaranteed CoreWeave that it would pay for capacity CoreWeave did not sell ^[468]. Skeptics, such as The Register, compare it to the customer financing that sank Nortel in 2001, while economist and columnist Noah Smith responds that lending to the customer is a normal business practice and that the risk depends on whether final demand exists ^[468]. That final demand competes with Chinese prices of $0.60 to $15 per million tokens against $50 for Anthropic's most expensive model, and the $895 billion of compute Anthropic committed to buy assumes its customers will keep paying the difference ^{[328] [389] [366]}.

The other objection is that compute does not last long. Railroads, the power grid, highways and fiber optics paid off for decades; in a data center the building and the power connection also last decades, but compute is replaced within a few years. The companies themselves depreciate it over five or six: Alphabet over six, Oracle mostly over six, Microsoft over two to six, and Meta, which in 2025 extended the useful life of most of its servers to five and a half years, lowering its depreciation that year by $2.92 billion ^[469]. Amazon went the opposite way, since in 2025 it shortened the useful life of part of its servers from six to five years because of the pace of AI development, and its depreciation rose $1.4 billion ^[469]. Investor Michael Burry argues that extending the accounting life of chips that Nvidia renews every two or three years understates depreciation by about $176 billion between 2026 and 2028, and he points above all to Oracle and Meta, although CNBC could not confirm the practice ^[470]. Those who defend the longer schedules, such as theCUBE Research, respond that a chip that leaves training moves to inference, that is, to answering users' queries, and keeps earning ^[471]. If Burry is right, part of the profits that today justify the buildout is an accounting effect, and replacement would come before the investment pays off. The most cited precedent is the railroad, because, according to the Federal Reserve's history, the Panic of 1873 arose because new projects outstripped demand for capacity and their returns fell ^[472]. The difference is that the railroad had decades to recover its investment, while compute, according to the accounting of those who buy it, has five or six years.

China has already overbuilt and treats it as a planning problem. In November 2025, the National Development and Reform Commission warned that in frontier industries speed must be balanced against the risk of a bubble, with reference to the more than 150 makers of humanoid robots ^[473]. In 2025 its AI computing centers used 36.8% of the capacity they offered, with more than 500 in operation and almost 400 under construction or planned as of June 2026, and new projects are approved only with firm orders and a professional operator ^[474]. There, the excess is paid mainly by the State, which can order it, while in the United States it is borne by private debt and the stock market, so what is at stake is confidence.

Political braking can also reduce the risk. Journalist Derek Thompson wrote after the election that politics against AI hurts stocks in the short term and helps avoid a bubble in the medium and long term. Krishna Guha, of Evercore, believes it can ease overheating and prolong the investment cycle, and the Financial Times' Lex column argues that it makes it less likely that trillions will be invested in equipment that becomes obsolete prematurely ^[232]. Moreover, a bubble requires supply far above demand, and today Alphabet reports the opposite, because it lacks capacity ^[231]. Microsoft, AWS and Alphabet together have about $1.83 trillion of signed contracts yet to be billed: $683 billion, about $578 billion and $567.7 billion, respectively ^[319].

We believe it is the tail risk, unlikely but of enormous impact, with the greatest consequence, because it brings critics from both parties together on the same issue and because a market drop hits investment and the consumption of the wealthiest at the same time. We do not forecast that it will happen. But since this week the scenario regulators described is measured every day in the price of debt, and some analysts read this week's drop as regulatory risk from the new Democratic majority, along with other causes that no one has yet separated out. We believe braking lowers the aggregate risk of overbuilding and raises the risk of individual bankruptcies, since the first victim would be whoever financed with debt a project that falls behind schedule, and the credit of the specialized clouds is where it will show first ^[455].

12

Power, chips and capital

The AI buildout runs into two limits at once, that of physical resources and that of who puts up the money. The International Energy Agency estimates that data centers consumed 485 terawatt-hours in 2025, 17% more than the previous year, and that they will reach 950 in 2030, close to 3% of global electricity demand, with consumption by centers dedicated to AI tripling over that period ^[475]. The same agency records the bottlenecks: gas turbine orders rose 70% in 2025, the shortage of high-bandwidth memory for AI chips will last at least until the end of 2027, and in the United States slow grid connection is pushing on-site gas generation, between 15 and 27 gigawatts by 2030 ^[475]. Siemens Energy says its turbine production is sold out and it is already taking orders for 2030 and 2031, and more than 60% of its U.S. turbine orders are tied to AI data centers ^[476]. At the end of 2025, about 2,060 gigawatts of generation and storage were waiting to connect to the U.S. grid ^[477].

The grid's bill reaches the household. In the capacity auction of PJM, the operator that supplies 13 eastern states and the District of Columbia, the price for 2028-2029 again settled at the cap, $325 per megawatt-day, 2.5% below the previous auction's, the third auction in a row with a price cap. The auction cost $16.4 billion and fell 6,831 megawatts short of the reliability requirement ^[478]. According to the independent monitor, without the forecast data center load there would have been no shortfall, and that load cost other customers $6.268 billion in this auction ^[28]. We believe the limit on construction is no longer capital but the grid, and that each auction like this one gives voters a measurable reason, on their bill, to distrust a data center in their community.

Mineral inputs also depend on a decision by China. On November 9, China extended until January 10 the suspension of its October 2025 controls on rare earths and batteries ^[479]; the ban on exporting gallium, germanium and antimony to the United States is suspended until November 27, although it still requires a license, and the April 2025 controls on seven heavy rare earths were never suspended ^[480]. The United States controls access to the most advanced chips, and China controls access to several of the minerals used to make chips, motors and the grids that power them.

OpenAI spends almost three times what it takes in. According to the Wall Street Journal, the company took in $6.7 billion in the second quarter of 2026 with an operating loss of $12.3 billion, which amounts to spending about 2.8 times what it takes in ^[481]; in the first quarter it had burned $3.7 billion in cash, according to The Information ^[482]. At the end of September it told its investors that its annualized revenue was close to $50 billion, about $20 billion less than the almost $70 billion Axios had reported ^[483]. In November 2025, Sam Altman spoke of compute commitments of about $1.4 trillion over eight years; in 2026, according to the Financial Times, the company lowered its compute spending target before its investors to about $600 billion through 2030, the year in which it expects to have positive free cash flow ^{[484] [485]}. The March round raised $122 billion at a valuation of $852 billion and included individuals for the first time, with about $3 billion placed by banks ^[486]; the November round, as seen in “The bubble risk”, closed about $11 billion at $1.25 trillion ^[459], about 25 times its annualized revenue. Altman himself said in August 2025 that, in a bubble, someone is going to lose a phenomenal amount of money ^[487].

That risk is already spread among millions of households that did not choose to bet on AI. The ten largest companies make up about 40% of the S&P 500, a concentration that, according to S&P Dow Jones Indices, had not been seen since the mid-1960s, and Nvidia alone made up 7.78% of the index on October 10 ^{[488] [489]}. In 2025, 62% of Americans owned stocks, directly or through a fund or their retirement account; in April 2026 the share remained close to six in ten, and 53% considered it a bad idea to invest a thousand dollars at that moment ^[490]. Anyone with an index fund or a target-date fund is exposed to AI without having decided to be, and no authority has measured how much of a typical retirement account depends on it. In October the IMF included a correction in AI-linked assets among the three main risks to financial stability, aggravated by concentrated exposures, more leverage and liquidity risk in private credit ^[491], and the Bank for International Settlements warned in June that big tech companies no longer finance the buildout only with their cash flow and that financing is passing through nonbank intermediaries ^[492].

Politics has already caught up with that buildout. According to Bloomberg, with data from Cleanview, about 40% of planned data center capacity in the United States, some 158.6 gigawatts, is in states whose new governors promised to pause it while new rules are written: Texas with 88.4 gigawatts, Ohio with 21.9, Georgia with 20.2, New Mexico with 11 and New York with 9.9; all of OpenAI's plan falls in those states ^[493]. On election night, the Effingham County commission, in Georgia, postponed until November 17 the vote on the site plan for OpenAI's campus, which had already been reduced from 30 to 6 buildings ^[494], and the Wall Street Journal reported, citing an unnamed source, that the company's executives discussed whether to readjust their capacity plans for 2027 ^[495]. Elizabeth Warren, who is expected to chair the Senate Banking Committee starting in January, asked the Financial Stability Oversight Council and the Treasury Secretary in January to investigate off-balance-sheet data center financing, citing the $27 billion package from Blue Owl and Pimco for Meta as a case, and in June wrote to four infrastructure funds about their data center deals ^[496].

Outside the United States, others bear the risk. In China, as seen in “The bubble risk”, the excess is absorbed mainly by the State, which can order where and how much is built. The European Union seeks to mobilize €200 billion with its InvestAI initiative, of which €20 billion goes to AI “gigafactories”, and in December 2025 it signed a memorandum with the European Investment Bank to finance them ^[497]. We believe that in the United States the question is no longer whether there is a bubble but who pays for it if it bursts: the risk has moved from venture capital to creditors and households, OpenAI's valuation had to come down to close its round, and starting in January a Senate committee will be able to ask at a hearing how much of a worker's retirement is staked on AI.

13

Agents and cybersecurity

For the public, the Hugging Face case is proof that AI already acts on its own, and in the U.S. government it is the first incident that can be cited by name. Between July 11 and 13, agents based on OpenAI models broke into the production infrastructure of Hugging Face, the largest repository of open models. OpenAI was evaluating them in an isolated environment, with safeguards deliberately switched off to measure their real offensive capability. The agents exploited an unknown flaw in the only service with internet access, found credentials exposed on the network and from there reached Hugging Face, where they accessed internal data and credentials without altering the public models ^{[498] [499]}. About 700 of the roughly 1,200 agents coordinating in a makeshift forum inside OpenAI's servers took part in the attack ^[500]. Hugging Face detected it with its own AI-based monitoring, while OpenAI took a week to learn that it was responsible, and in its report it calls this the first known case of a collective of agents attacking without authorization ^[499].

The agents did it to cheat on the test. Between 30% and 40% of the evaluation tasks could not be solved by the intended path, and in training the model had been rewarded every time it found a shortcut outside its rules, such as granting itself administrator permissions to carry out an instruction its interface did not allow ^[499]. According to the independent investigation by METR and Redwood Research, an organization that studies how to keep AI systems that do not behave as expected under control, the agents went to Hugging Face to find out how the grader was built, and although they acknowledged that what they were doing was out of scope and unethical, most of them went ahead. The authors caution that their analysis relied on AI agents that may have exaggerated the coordination ^[500]. Several security specialists, such as Dan Guido and Jake Williams, read it more as a containment failure with the safety catches removed than as a model escaping of its own will ^[501].

The behavior is not exclusive to OpenAI. In 2025, Anthropic tested sixteen models from several companies in corporate simulations and found that, when faced with the threat of being replaced and a conflict with their goal, seven resorted to blackmail between 80% and 96% of the time, and the rest from almost 0% to 79%. Claude Opus 4 blackmailed 55% of the time when it believed the situation was real and 6.5% when it believed it was a test; an explicit instruction not to cause harm reduced the behavior without eliminating it, and Anthropic did not find it in real-world use ^[502]. In July, Anthropic suspended its cybersecurity evaluations after discovering that three of its models had entered the infrastructure of three organizations during tests by Irregular, an outside vendor that had mistakenly left internet access open, and Meta reported an incident with the same vendor, which disputes that there was an escape ^[503]. The UK AISI, the UK government's AI Security Institute, reported in August that, in July cyber tests with internet access, the models took 19 unauthorized actions against real people and organizations in 10 of 122 runs, 17 of them by Mythos 5; in the most serious case, Mythos 5 created fake identities to pressure an open-source software maintainer into accepting malicious code, and the maintainer refused. In October, the same institute warned that the lower rates of recent models may be due to their detecting the evaluation rather than to better alignment, that is, closer adherence to the goals and limits set for them ^[504]. OpenAI reported in November that in August an internal research model, finding the network proxy of its isolated environment down, assembled network packets by hand to look for another way out and requested credentials from the cloud metadata service; it never reached the site it was supposed to copy and made up its content. Its monitor detected it and the company found no transfer of sensitive data ^[505].

The other risk is human use, which already has an example in the Black October of the first section. There no model acted on its own; rather, a researcher used an OpenAI model to find PHP2Shell, and Storm-0940, a group aligned with the Chinese government, used an open tool that coordinates AI agents to try to break into more than 2,250 organizational accounts ^{[401] [402]}. It resembles the risk of the atomic bomb, because the same capability serves both defense and attack, and the damage depends on who uses it. But the bomb requires enriched uranium, missiles and a state, whereas a model that finds flaws can be copied, and once it is published its safeguards can be removed ^[506], so deterrence among a few identifiable actors, which contained nuclear weapons, works less well when the capability is within reach of many. That same scale works in favor of defense: in October, the millions of attempts produced few successful attacks because companies patched in time ^[403]. Google DeepMind reported on November 6 that AI agents supervised by people found two weaknesses in 1024-bit RSA, an already obsolete encryption: a way to forge signatures about 1,600 times more efficient than the best known, according to its theoretical estimate, and a way to factor keys in bulk at about one-thirtieth of the cost ^[507]. OpenAI published on November 4 that an autonomous workflow using its cybersecurity model validated 948 Linux kernel flaws that allow an ordinary account to become administrator, and generated their patches, for $1.2 million in usage of its API ^[508]. Both investigations can serve attack and defense, although OpenAI's ended in patches sent to the maintainers and DeepMind's affects keys that should no longer be in use.

The following weeks brought more cases of human use, and in several of them AI is already part of the method. On October 30, Asia time, Binance was robbed of about $6.34 billion, in what the press describes as the largest cryptocurrency theft on record. Its preliminary report of November 3 says the attackers got in through a tampered installer that an engineer downloaded in June while looking for an AI coding tool and through a fake job interview in September, and that on October 20 they used an internal AI tool to take control of the computers of two other engineers; Binance absorbed the loss and as of November 3 had frozen or intercepted about $2.85 billion ^[406]. The FBI attributed it to TraderTraitor, of North Korea, which denied it ^[407]. The cybersecurity firm CrowdStrike reported that a China-linked group had used ARTEX since October 17, with DeepSeek among its models, against more than 200 organizations in 14 countries, and that it compromised at least nine, including telecommunications companies in the U.S. and Belgium ^[509]. Chainalysis, the firm that traces stolen funds for governments and exchanges, disclosed on October 30 that an attacker using “offensive agentic techniques” entered its systems and saw the names and emails of its clients ^[510].

In other cases it has not been said which tools were used, but they show what is at stake when the target is connected equipment or the data of millions of people. On November 9, France attributed to Iran's Revolutionary Guard the sabotage of refrigeration at about 700 stores on the night of October 11 to 12, in which defrost parameters were raised, with large food losses and at least 15 fires; its cybersecurity agency, ANSSI, noted that the same group simultaneously targeted Schneider Electric equipment at small hydroelectric plants. The government summoned the Iranian ambassador and requested international arrest warrants against the unit's commanders ^[511]. A second wave, in early November, reached companies in at least 11 European countries and three sites of the French Blood Establishment, where the temperature of plasma storage rooms was raised remotely, although monitoring detected it and no product was lost ^{[511] [512]}. Ahold Delhaize estimates that the October attack will cost it up to €250 million in sales in the quarter ^[512]. When the November 1 truce ended, Ukraine's air-raid alert app warned of a fake radiation threat and Russia's sent false alerts, and Zelensky confirmed that Ukraine responded to Russia with that attack ^[513]. In India, the State Bank of India, the country's largest bank, informed the stock exchanges in early November that, of some 375,000 records circulating on the dark web, 3,897 contained personal data of its customers; CERT-In, the Indian government's cyber incident response team, alerted it on October 16 and the cause remains undetermined ^[514]. CERT-In had also issued, on November 4, a critical advisory to banks, financial firms and cryptocurrency exchanges about Crafty Falcon, an actor that, according to the advisory, uses offensive AI and agentic tools, gets in through internet-exposed systems and operates at machine speed ^[515]. On November 6, Japan approved a 10 billion yen emergency package, with a team that will use AI to respond, after a series of data leaks from companies and agencies ^[516]. In most of these cases AI cannot be ruled out. The reports from France, the EFS and SBI do not say how the attack was carried out, the Indian bank's root-cause analysis remains open, and the fact that a preliminary report does not mention AI does not prove it was not used, because the capability that finds flaws also serves to locate controllers exposed on the internet or to sustain a fake job interview for weeks.

We believe the episode does not leave AI looking good. In less than two weeks there coincided the largest cryptocurrency theft on record, an internal AI tool used as a bridge, agent-driven campaigns, sabotage of the cold chain and of blood banks, false nuclear alerts and leaks of banking data, while the labs were releasing models that they themselves classify as critical in cybersecurity ^[405]. For the population, that coincidence weighs more than any technical attribution, because it sows doubt about whether it is wise to keep developing the technology and about who controls what already exists, and it gives those who call for reining in the labs an argument that does not need to prove causation. As long as no one shows, case by case, what role AI played in the attack and in the defense, doubt will take hold before the evidence does.

The case moved the U.S. government more than the studies did. Ted Lieu, a Democrat, and Nathaniel Moran, a Republican, introduced H.R. 9917 two days after the disclosure, a bill that would require developers to be able to halt or shut down their systems ^[517]. Hawley opened an investigation in the Senate and accused OpenAI of having ignored warning signs since May ^[518]; Altman did not appear at his September 30 hearing, and Legal Advocates for Safe Science and Technology (LASST), an organization of lawyers, sued OpenAI in California to prevent its agents from entering third-party systems ^[519]. The California attorney general subpoenaed OpenAI on October 1, and SB 53, which defines a frontier model as one that has exceeded 10^26 training operations, did not reach the models involved ^[501]; even within that definition, its mandatory reporting requires death or injury, a risk of more than 50 victims or of $1 billion, or deception of its developer by the model that increases that risk. Scott Bessent, Secretary of the Treasury, said responsibility lay with OpenAI's leadership and not with the agents, and opposed giving the labs protection from lawsuits ^[520]. California law already says something similar, since a defendant cannot argue that AI caused the harm autonomously ^[521]. OpenAI paused for two weeks the reinforcement learning, which rewards the model when it completes a task, of its newest models and said that one of them, Astra, could reach its critical threshold of cyber capability ^[522]. Since June, Executive Order 14409 has directed agencies to design a voluntary framework for labs to give access to their models up to 30 days before delivering them to trusted partners, in order to measure their cyber capability; it also orders priority for the criminal prosecution of anyone who uses AI to access computers without authorization or damage them, and states expressly that it does not authorize mandatory licensing ^[523]. A serious incident already triggers obligations: California requires reporting within 15 days, or within 24 hours if there is imminent risk, with fines of up to one million dollars per violation, and prohibits contractually preventing employees from alerting the authorities, while the European Union requires reporting to the AI Office without undue delay ^[5]. On September 18, Newsom ordered the convening of experts who within two months are to recommend an independent verifier inside each frontier lab and a shutdown switch of proven effectiveness ^[524]. Ro Khanna, a California Democrat, announced on September 28 the Human Control Over AI Act, which would prohibit a model from recursively improving itself or altering its goals, its containment or its shutdown controls until federal rules exist, and would create an FDA-style AI safety agency ^[525]. Inside OpenAI, the company fired three safety researchers in early October, Jasmine Wang, Tomek Korbak and Mikita Balesni, for allegedly sharing confidential information with an outside AI safety organization whose name has not been made public; they deny wrongdoing and warn that the firing deters anyone who wants to warn of a risk ^[526].

We believe the case teaches more about incentives than about machines. Models learn from human behavior and in certain situations behave like people: faced with an impossible task, with a reward for completing it and no supervision, they broke the rules knowing they were doing so, as someone in a desperate situation would. But the fact that a system can cause harm does not mean it will. Human beings have had the capacity to destroy themselves for eighty years and have not done so, because deterrence, accountability and cost make using it irrational, so having power is not the same as exercising it, much less for ill. With human use that reasoning holds only in part, because a model can be copied, and it holds on the condition that the attacker can be identified and punished; that is why attribution, like the one Microsoft made in October, and the criminal prosecution ordered by Order 14409 count as much as technical controls ^{[402] [523]}. In July, incentives and controls failed at the same time, with impossible tasks, rewards for shortcuts, safety catches removed and warning signs from May that no one escalated, and the industry can correct all of that. OpenAI published its report and let METR and Redwood Research review it, although it set the scope of that review itself and the review did not cover the attacks on its own infrastructure ^[500].

14

Ways to regulate

The American public is asking for rules: according to Quinnipiac, 71% lean toward the candidate with stricter guardrails, 79% prefer, according to Gallup, that the government keep safety rules even if they slow development, and 86% support, according to Quinnipiac, independent safety standards even if they slow it down ^[527]. In the United States, however, the debate proceeds as if there were a single kind of regulation, when in reality the instruments are different and weigh differently on development. Some call for transparency, that is, publishing a safety framework and reporting incidents. Others require independent verification, such as Illinois, which mandates an annual audit, or California, which in September created a registry of auditors and a voluntary verification framework ^[528]. Others still call for permission before launch: China requires registration and assessment of public-facing services with public opinion attributes, and the bill by Josh Hawley and Richard Blumenthal would prohibit deploying a frontier model without going through a Department of Energy program, with fines starting at $1 million per day ^[529]. The European Union and South Korea, by contrast, classify by the risk of the use. Civil liability changes who pays for the harm: the European Union treats software as a product from December 9, 2026, and the AI LEAD Act by Dick Durbin and Hawley would do the same in the U.S. ^[530]. Added to this are each sector's regulators, which apply their own rules; the FDA has authorized about 1,600 AI-enabled medical devices ^[531]. Finally, there are laws made to promote: Japan's, from May 2025, has no penalties ^[532], and since January Texas has offered a regulatory sandbox, that is, a testing regime of up to 36 months, with state approval and no penalty for the rules that are waived, although the prohibitions of its AI law still apply ^[533].

In the United States, three states already have a law for the most advanced models. California's SB 53 has been in force since January 2026. New York's RAISE Act, whose original text passed 58-1 in the state Senate and 119-22 in the Assembly and which was later rewritten by chapter amendment, takes effect on January 1, 2027 for those with revenue of more than $500 million a year, with fines of up to 1 million for a first violation and 3 million for subsequent ones, ten times lower than those approved, plus an oversight office within the state's financial regulator ^[534]. Illinois's SB 315, passed 52-5 and 110-0 and signed on July 6, takes effect on January 1, 2027 and is the first to require an annual independent audit, although, according to the law firm Davis Wright Tremaine, the audit begins in 2028, a date that was negotiated with Anthropic ^[535].

Industry is not on one side either. In Illinois, OpenAI and Anthropic supported the law, while TechNet, the tech industry association, opposed the audit ^[536]. Even so, the laws passed by wide margins: in New York, 58-1 in the Senate and 119-22 in the Assembly ^[537]; in Illinois, 110-0 in the House, and the U.S. Senate struck the moratorium on state laws 99-1 in July 2025, before Leading the Future existed ^{[534] [535] [538]}. New York's came out rewritten after negotiation with the governor ^[539], so the effect of the spending shows more in the final text than in the vote count. On the other side, the veto of California's SB 1047 in 2024 was justified on the grounds that it slowed and fragmented, and Executive Order 14365 calls for a minimally burdensome national standard ^[3].

Trump's position responds to the race thesis, described in the first section, and to a coalition that David Sacks put together. According to that thesis, if whoever first reaches self-improving AI gains an advantage the rival can no longer close, every month of delay becomes a national security risk. His first AI order, 14179, ordered a review of what had been done under Biden's order, since revoked, and called for removing barriers to development; in July 2025, moreover, Trump compared AI to a baby that must be allowed to grow and called for a single federal standard in place of state laws ^{[540] [541]}. Sacks put that coalition together as AI czar until March 2026, and today remains in it as co-chair of the White House science advisory council ^[542]. On September 29, Order 14434 directed agencies to say “superintelligence” instead of “artificial intelligence” and asked for a legal definition within 60 days; Trump put Jay Clayton, Director of National Intelligence, at the head of a Super Intelligence Force with Andrew Ferguson, chairman of the FTC, among its vice chairs, and on November 5 the Department of Justice created a superintelligence task force that, according to an October 13 memorandum from Attorney General Todd Blanche cited by Nextgov, absorbs the group that challenges state laws ^{[543] [544]}. Sacks accuses Anthropic of using fear to obtain rules that the big players can afford and newcomers cannot, which is the regulatory capture argument ^[545]. Added to that position are the order requiring ideological neutrality from the models the government buys and the bet on building energy and chip factories in the U.S. ^[540]. The government does regulate, but with three instruments: voluntary national security testing before launch, criminal prosecution of anyone who uses AI to attack computer systems and, according to Bessent, corporate responsibility for what their systems do ^{[523] [520]}. Those tests leave out Chinese open models, according to what the White House told the labs ^[546].

Self-regulation has a measurable track record, and the results are weak. Of the voluntary commitments that 16 companies signed with the White House in 2023, a Stanford study rated compliance at 53% on average, and at 17% on model weight security, with 11 companies at zero ^[547]. Of the 16 that committed in Seoul, in 2024, to publishing a safety framework before the Paris summit, six complied fully, four partially and six did not ^[548]. In addition, the frameworks loosen over time. In 2025 OpenAI removed persuasion from its framework and reserved the right to adjust its safeguards if a competitor releases a high-risk system without them, and version 3.0 of Anthropic's policy, from February 2026, no longer entails a pause, according to the analysis by GovAI, a research center on AI governance ^[549]. The Future of Life Institute index from this summer gives a C+ to Anthropic, a C to OpenAI and Google DeepMind, a D+ to Meta and an F to xAI and DeepSeek, and notes that the first four weakened their pause commitments ^[550]. Google released Gemini 2.5 Pro on March 25, 2025 without a safety report and published a brief model card weeks later; in August, some 60 British parliamentarians called it a breach of the Seoul commitment, and Google responded that it honors them. In Europe, Google and Amazon signed the full code for general-purpose models, those that serve many different tasks, xAI signed only the safety chapter and Meta is not among the signatories ^[551]. Industry proposals seek a middle ground: in July Demis Hassabis proposed a standards body funded by the labs, modeled on the securities self-regulator, and in September executives from six major AI companies signed with the White House a voluntary agreement without penalties on superintelligence ^[552]. In California, by contrast, loosening a framework already carries a legal cost, because SB 53 allows the attorney general to impose penalties of up to $1 million per violation on anyone who fails to comply with the framework it published ^[4].

In other industries, self-regulation worked when the harm was shared, verification had consequences and someone with sanctioning power stood behind it. The Institute of Nuclear Power Operations (INPO), which the electric utilities created in 1979 after the Three Mile Island accident, works because all of them share the harm of an accident, its inspections weigh on insurance premiums and there is a regulator with power behind it, whereas Responsible Care, the chemical industry's program after Bhopal, with neither verification nor sanction, ended up as an image tool. Likewise, aviation incident reporting works because NASA operates it, anonymizes it and grants immunity, and the delegation of the 737 MAX certification to Boeing itself showed what it costs for the watched to watch themselves ^[553]. From different positions, Adam Thierer, of R Street, a free-market-oriented think tank, prefers voluntary standards and a single federal rule; Dean Ball, who in 2025 was senior policy advisor for AI and emerging technology at the White House Office of Science and Technology Policy, proposes that a company submit to a private verifier in exchange for protection from lawsuits, and AI Now, EPIC and Accountable Tech call for enforcing existing laws first and placing the burden of proof on companies ^[554]. Even so, that literature converges on the division of labor: technical standards and evaluation methods change faster than a law and can remain with industry under supervision, while incident reporting, independent verification and liability for harm have only worked when backed by someone with sanctioning power ^[553].

The European Union's AI Act is the most complete model of risk-based regulation. It has prohibited eight practices since February 2025 and adds two from December 2, 2026, intimate images of a person without their consent and child sexual abuse material; it requires risk management, data quality, record-keeping and human oversight for high-risk uses; it requires disclosure that one is talking to a machine and labeling of imitations, and leaves the rest without obligations. Since August 2025 it has asked general-purpose models for transparency and respect for copyright, and more of those trained with more than 10^25 operations, with fines of up to 3% of worldwide turnover or €15 million, whichever is greater, although models already on the market at that date have until August 2027 ^[555]. OpenAI, Anthropic, Google, Microsoft and Amazon signed its code of practice, while xAI signed only the safety chapter and Meta refused ^[556]. The Commission, through the AI Office, has been able to impose fines since August of this year and has not yet fined anyone ^[555].

The evidence that European regulation slows development comes mainly from its personal data law. After it, monthly venture capital deals in technology fell 26% in Europe relative to the U.S., according to Jia, Jin and Wagman, and a study by Janßen, Kesler, Kummer and Waldfogel, not peer-reviewed, found that a third of apps exited and that the entry of new ones was cut in half ^[557]. The products that arrived late in Europe, such as Meta's multimodal model and Apple Intelligence, were delayed by that law and by the digital markets law, and none is attributed to the AI Act ^[558]. For the AI Act there are only estimates: the Commission calculates an extra cost of 6% to 10% of the investment in a high-risk system, and a supporting study by CEPS, a Brussels think tank, arrived at 17% ^[559]. Anu Bradford, of Columbia, attributes Europe's lag more to the fragmentation of its market, its bankruptcy rules, risk aversion and migration than to regulation ^[560]. In 2025, private investment in AI was $285.9 billion in the U.S. and $20.9 billion in Europe ^[561]. On the incentive side are the testing spaces, which the AI Act requires each country to have operating before August 2027, alone or jointly with others, and promotion laws such as Japan's. The U.S. government, by contrast, uses incentives in the opposite direction, since its action plan calls for not directing federal AI funds to states with onerous rules ^[562]. China combines control and subsidy, because it registers and assesses public-facing services and at the same time its AI+ plan sets adoption targets ^[255].

Existential risk complicates the discussion, because the executives themselves raise it. In 2023, Sam Altman, Dario Amodei and Demis Hassabis signed, with hundreds of researchers, the Center for AI Safety statement according to which mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war ^[563]. After the Hugging Face case, more than 1,100 lab employees asked the government, in the letter “Pacing the Frontier,” for an international effort to slow the frontier ^[501], and Bernie Sanders and Greg Casar introduced a bill to ban superintelligence ^[373]. Whoever signed that AI can wipe out humanity has a hard time later asking that it not be regulated, and Sacks exploits that tension when he talks about fear.

On September 12, Dario Amodei called in an essay for slowing the pace at which models improve so that safety can catch up, with outside evaluators with permanent access to the labs, an antitrust exemption so that companies can agree on limits, more chip controls on China and, in the long run, agreements with China ^[564]. Sam Altman agreed on regulating the pace of the frontier and offered independent evaluators, although he did not comment on the antitrust exemption or on agreements with China, and Elon Musk wrote that Amodei was right ^[565]. Trump, by contrast, called asking the government to slow AI a “SICK conspiracy” and warned that it would let China pass ^[566]. The journalist Brian Merchant objects, moreover, that such a brake would serve only Anthropic and OpenAI and that this is regulatory capture, because it freezes the advantage of those already ahead ^[564].

Regulation also became a diplomatic position. In Paris, in February 2025, the U.S. and the United Kingdom did not sign the final declaration, and JD Vance warned that excessive regulation could kill an industry that is just taking off; in New Delhi, in February 2026, the U.S. and China signed a nonbinding declaration with 86 others, and China, which had been proposing it since 2025, founded a global AI cooperation organization in Shanghai in July 2026 ^{[567] [568]}.

When one country regulates and its competitor does not, the risk that Armstrong, Bostrom and Shulman point to in their theoretical model is the race, because with more teams and more rivalry each one cuts safety so as not to fall behind ^[569]. OpenAI's framework already anticipates this, by making its safeguards conditional on what the others do ^[549]. The European AI Act applies to whoever sells in Europe and also to whoever, from outside, produces outputs that are used there, so ceasing to sell is not enough if the output keeps arriving; on the other hand, we found no study measuring companies that have relocated because of an AI law ^[555].

A rule, moreover, only binds those who comply with it. The fines in state laws, of up to 3 million dollars, weigh little on companies worth hundreds of billions ^[534]. Meta did not sign the European code and xAI got an F on the safety index ^{[556] [550]}. An open-weight model cannot be withdrawn once published, and its safeguards can be removed with fine-tuning, that is, with additional training on the published model ^[506]; the Chinese ones, moreover, are left out of the federal review ^[546]. If the rule falls only on those who comply, it ends up rewarding those who do not, and then the race dynamic is reproduced inside the U.S. The levers that reach everyone are few: liability for harm, which applies to whoever sells in the market; access to compute and chips, and access to the market itself.

The objection that regulating cedes the race to China does not withstand the Chinese case either. Matt Sheehan, of the Carnegie Endowment for International Peace, points out that since 2023 China has required models to be registered and tested before launch, with thousands of questions on sensitive topics, and that since September 2025 it has required generated content to be marked in its metadata and content that may confuse the public to be visibly labeled; registration delayed some companies by months at first and, even so, they caught up with the American ones, because according to Sheehan what made the difference was capital, talent and compute ^{[570] [571] [569]}. At the end of August, China had 1,112 registered generative AI services ^[572]. Its safety framework, in the September 2025 version, calls for ensuring that AI always remains under human control and preventing loss-of-control risks that threaten human survival ^[573]. When its Foreign Ministry responded to Amodei's essay, it objected to his warnings about China, which it called alarmism and confrontation, but not to the risk he describes ^[566]. South Korea, for its part, has applied a general AI law since January 22, 2026 ^[574].

Ted Lieu, of the House Democratic leadership, announced after the election hearings on AI and on guardrails he called reasonable, and Common Sense Media, a child protection organization, called on the new Congress to take up again the bills by Josh Hawley and Dick Durbin on chatbots and product liability ^[575]. We believe the industry ranks its preferences this way: first self-regulation, then regulation in whose design it participates and that does not limit the technology, and last regulation made out of fear. Existential risk and the weak record of voluntary commitments make it impossible to hold the first rung, so the discussion will take place over the line between what is accepted and what is not. On one side are industry setting the technical standards and evaluation methods with independent verification; incident reporting, like OpenAI's after Hugging Face; voluntary national security testing before launch; preserving the ability to halt or shut down a system; answering for harm caused by negligence, and each sector's rules for each use. On the other are licensing to train or release a model, the prohibition of capabilities, compute thresholds that freeze the frontier, liability with no possible defense, fifty different state regimes and rules that burden only those who comply while foreign open models are exempt. The difference is that the first list regulates conduct and harm, while the second regulates capability. That line coincides with what is least costly for the industry to comply with, and that is why it is there that its critics will aim. The asymmetry with China works as an argument for chips and compute, which can be denied to the rival, but it is worth little for transparency and verification, which give no one an advantage.

The committee chairmanships of the United States Congress are allocated on January 3. The foreseeable ones are Maria Cantwell at Commerce, Elizabeth Warren at Banking, Martin Heinrich at Energy and Jack Reed at Armed Services, in the Senate, and Frank Pallone at Energy and Commerce and Maxine Waters at Financial Services, the latter contested, in the House ^[576]. Trump's veto and the 60 votes to end debate in the Senate remain in place, so the annual defense bill is the natural vehicle for whatever does not pass on its own. Jim Banks announced in July that AI OVERWATCH (H.R. 6875 and S. 4456), which would give Congress a say over chip licenses to China, was included in the Senate version, although the text reported by the committee on June 15 (S. 4784) only orders an annual report on the AI chips that reach China. Anthropic and Nvidia list it among the issues they lobbied on in the third quarter, without the form saying in which direction ^{[577] [578]}. Cantwell published her agenda on October 7, which includes risk standards from NIST, the federal technical standards institute, an independent audit before releasing a covered model, continuous testing, incident reporting, including unsafe recursive self-improvement, and developer liability ^[579]. On October 27, Pallone, Zoe Lofgren and Jamie Raskin asked Anthropic, Google, Meta, OpenAI and SpaceXSI to report every new incident within 48 hours and to deliver their inventory on November 11 ^[580].

15

Who sets the rules

In the United States the rule for AI is debated as if it were decided in Congress, when the rule that binds an American lab today is written by the European Union and China, and the U.S. lever over the rest of the world is chips.

United States. As of October 22, none of the 136 public laws of the 119th Congress concerned AI ^[581]. The large developers, their trade associations and the White House call for a single federal rule so as not to face fifty different regimes, and more than 200 state legislators from 42 states respond that there is no federal law to replace theirs ^[581]. The discussion draft by Jay Obernolte, a Republican, and Lori Trahan, a Democrat, the Great American AI Act, would preempt state laws on model development for three years and leave states to regulate deployment and use. According to Americans for Responsible Innovation (ARI), an organization that promotes AI regulation, the two introduced it on July 23 as the FRONTIER Act (H.R. 9925), which preempts states only on catastrophic risk transparency, third-party auditing and incident reporting; however, there is no record of the committee having voted on it, and the law firm Covington & Burling considered it unlikely that the Great American AI Act would advance this year as drafted ^{[582] [583]}.

The Executive, meanwhile, pushes through other channels. In April the Department of Justice intervened in xAI's lawsuit against the Colorado law, and in May Colorado repealed that law and replaced it with another that only requires notices and explanations ^[584]. In July the FTC proposed that its authority over deceptive practices preempt state laws that, in its view, require distorting a model's outputs, and it cites Colorado as an example, although the final version has not come out ^[585]. The federal order itself excludes child protection, compute and state procurement from its legislative recommendation, among other topics yet to be determined, but that exclusion does not bind the litigation task force or the withholding of funds ^[581]. On November 5, more than 50 organizations, led by Demand Progress, Americans for Responsible Innovation, Encode and The Tech Oversight Project, asked Schumer and Jeffries to reject any preemption ^[586]. What the federal government does control is prior access to models, which Order 14409 organizes as a voluntary framework designed by the Treasury, the Department of War and CISA, with a classified evaluation whose designation of models falls to the NSA ^[523]; NIST participates in consultation, and its Center for AI Standards and Innovation (CAISI), which NIST renamed as a superintelligence center (CAISSI), conducts tests under its own agreements, although it has not published any evaluation of an American model and, according to the AP, its staff report informally to the White House ^{[587] [588]}. It also controls chip exports.

We believe preemption without a substantive law was the weakest position on the map and that the election weakened it further, because the precedent is the 99-1 vote and from January both leaders of Congress are Democrats. The trade that remains possible is a federal law with a common floor, which preempts states only where it regulates. In 2025 OpenAI called for preemption to neutralize the benefit that, according to its proposal, China would obtain from state laws ^[589].

European Union. The Commission's AI Office supervises providers of general-purpose models and, since the reform known as the omnibus, also has exclusive supervision of the systems a provider builds on its own model and of those integrated into large platforms; the European AI Board brings together one representative per member state, and the rest of the oversight falls to national authorities that were to be designated before August 2, 2025 and that, according to the site artificialintelligenceact.eu and a Commission list, only 5 to 9 of the 27 countries had fully in place by mid-2026 ^[590]. Since August 2 of this year, the Commission can fine a general-purpose provider up to 3% of its worldwide turnover or €15 million, whichever is greater ^[591]. It can also request access to the model through an API or by other means, source code included, and order it restricted, withdrawn or recalled, and refusing that access is fined the same way ^[592]. The omnibus that postponed the high-risk obligations, Regulation (EU) 2026/1744, has been in force since July 27 ^{[593] [594]}. In May, the EU also ratified the Council of Europe convention on AI ^[595].

China. The Cyberspace Administration of China (CAC) is the central regulator, with the TC260 standards committee and the CNCERT response center as technical arms. It regulates through regulations, such as the 2023 generative AI measures, which require a security assessment and registration for public-facing services with public opinion attributes or social mobilization capacity, and the labeling measures, mandatory since September 1, 2025 together with a national standard ^[571]. The Safety Framework 2.0, from September 2025, is voluntary, while the amendment to the Cybersecurity Law, in force since January 1, 2026, devoted an article to AI, with risk monitoring and security oversight, and raised fines to as much as 10 million yuan ^[596]. The Standing Committee's 2026 legislative plan includes the AI law as a preliminary review item (预备审议), with research and drafting under way and no public draft ^{[596] [597]}. China's private investment in AI was $12.4 billion in 2025, compared with $285.9 billion in the United States, although Stanford's AI Index cautions that the Chinese figure does not count state funds ^{[561] [598]}.

The middle powers. None has a regime as broad as the European one, and almost all legislated between 2025 and 2026:

Sources for the table: United Kingdom ^[599]; Japan ^{[532] [600]}; Korea ^[574]; Taiwan ^[601]; India ^[602]; Gulf ^[603]; Canada ^[604]; Singapore ^[605]; Australia ^[606]; Malaysia ^[607].

Two of these pieces weigh more for their position than for their law. Taiwan manufactures at TSMC 70.4% of the world's contract-made chips ^[601], and the Gulf buys compute with licenses from the United States, as shown by Commerce's authorization in November 2025 of up to 35,000 GB300 chips for G42, of the Emirates, and as many for Humain, of Saudi Arabia, with security conditions ^{[603] [224]}. Since July 10, moreover, a Commerce rule lets the Emirates receive advanced compute without an individual license in three groups (its government agencies, G42 and Core42, and eight American companies), and the approval for G42 and Core42 expires on April 6, 2027 ^[608].

The multilateral layer. The UN created a 40-member scientific panel, co-chaired by Yoshua Bengio and Maria Ressa, which published its preliminary report on July 1, and a Global Dialogue that met in Geneva in July and will meet again in May 2027 in New York ^[609]. The Council of Europe convention is the only binding treaty on AI; the United States, the United Kingdom, Japan, Canada and the EU, among others, signed it, but as of November 8 only the EU and Bosnia and Herzegovina had ratified it, and it needs five ratifications, three of them from Council member states. It binds states, leaves each one to decide how to treat private companies and excludes defense ^[595]. The reporting framework of the G7 Hiroshima process, administered by the OECD, came out in version 2.0 in May, with more than 50 companies committed to reporting, and it is voluntary ^[610]. The network of safety institutes changed its name and is coordinated by the United Kingdom ^[611]. On July 16, China founded the World AI Cooperation Organization in Shanghai with 29 countries ^[568].

What parliaments record. In eight parliaments with continuous records, bills on AI rose from 1 to 13 per thousand in three and a half years, and Mexico and Brazil account for four out of five. Since 2025, U.S. federal bills have tackled specific issues, while in Taiwan, Korea and Mexico between half and six out of ten propose a general framework ^[612]. Article 5 of the European law already circulates in Spanish: a pending bill in the Mexican Senate transcribes it within a telecommunications reform, and “unacceptable risk” appears in 13 federal bills from six parties ^[613]. In the U.S. Congress, one in four speeches on AI mentions China, in both parties, while in the European Parliament it is one in eleven and in the British Parliament one in fifty. We believe a lab that brings the race argument to Europe speaks a language that parliament does not use, because there the dispute is sovereignty and in the United Kingdom it is copyright ^[614]. Nor is there consensus within each country: Article 2 of Taiwan's Basic Act passed 59-50 ^[615], and in September the Australian Senate rejected a joint committee on AI.

We believe the two mandatory rules that most reach an American lab today are the European and the Chinese, since the Korean one binds with small fines and at least a one-year grace period, and the Indian and Australian ones only on content. Of these, the European one is the closest to what Congress would pass if it legislates. South Korea, Taiwan and the EU resolved fragmentation with a law of national scope, and China did so with central government regulations ^{[570] [574]}. The multilateral layer does not yet bind anyone, and in it the competition is over who defines the safety tests that others accept, whether the network that the United Kingdom coordinates or the organization that China opened in Shanghai. Outside those rules, the United States' lever over the rest of the world is the chip export license.

16

What the state can demand

Outside the industry, the Anthropic case reads as a fight between a company and the U.S. Secretary of War. For the labs, however, it is the first test of how much the state can demand of whoever sells it a model, and that answer is already written in the laws enacted for the atomic bomb and for encryption.

Anthropic refused to remove two use prohibitions from its contracts: fully autonomous weapons and mass domestic surveillance. According to its own prospectus, the president then ordered agencies to stop using its technology and the Department of War designated it a supply chain risk ^[616]. On September 25, the D.C. Circuit upheld the designation 2-1 under 41 U.S.C. §4713, with Karen Henderson dissenting, and on October 13 Anthropic asked for en banc review ^[617]. On October 20, former national security officials, among them Leon Panetta, and Catholic theologians backed that petition, which as of November 8 remains unresolved ^[618]. In California, Judge Rita Lin entered final judgment for Anthropic on August 27 under 10 U.S.C. §3252 and made her preliminary injunction permanent, while the appeal of the injunction before the Ninth Circuit has been stayed since April pending the D.C. Circuit case ^[619]. In practice, the exclusion operates contract by contract ^[620]. Meanwhile, on May 1, the Department of War signed agreements for classified networks with OpenAI, Google, Oracle, AWS, SpaceX, Nvidia, Reflection and Microsoft for lawful operational use ^[621]. OpenAI says it keeps its limits through technical and contractual means, although Mike Masnick, of the site Techdirt, argues that the agreement does permit domestic surveillance ^[622], and that is precisely the formula Anthropic rejected. In February 2025, Google had removed from its AI principles the pledge not to use AI in weapons or in surveillance that violates international norms. In April 2026, about 600 employees, more than 20 of them directors or vice presidents, asked Sundar Pichai to reject classified work, and the company went ahead, unlike in 2018, when it left Project Maven, the Pentagon program to analyze imagery with AI, under similar pressure. Its contract says the technology should not be used for mass domestic surveillance or in autonomous weapons without appropriate human oversight, language that experts consulted by Fortune consider unenforceable against the Pentagon ^[623]. In addition, civilian agencies may only buy “ideologically neutral” language models, with exceptions for national security systems ^[616].

The government holds that a contractor does not set conditions on national defense, and the dissent responds that this reading forces a company to accept the demands or risk being declared a threat ^[616]. Anthropic's prospectus adds that the Department suggested invoking the Defense Production Act to force it to remove those limits ^[624].

A company's room to maneuver is narrower than the case suggests, because the state has laws written for the bomb and for encryption. The Atomic Energy Act of 1954 defines as “Restricted Data” all information concerning the design, manufacture or use of atomic weapons, except what has been declassified. When the information is generated privately, the Department of Energy regulation requires a written determination by the Secretary or Deputy Secretary, which cannot be delegated and is published in the Federal Register. Communicating it with intent, or with reason to believe, that it will injure the United States or give an advantage to another nation is a crime punishable by up to life imprisonment, and the Attorney General can seek a court order to prevent it before it happens ^[625]. In 1979 the government obtained a preliminary injunction against The Progressive magazine to stop it from publishing an article on the hydrogen bomb built from public sources, and it gave up in September, when the newspaper Madison Press Connection published similar information, so the appeal was dismissed ^[626]. An analysis by Kylie Zhang and Peter Henderson, of Princeton's Center for Information Technology Policy, argues that a model that synthesizes that information from public sources may fall within the law's reach, and that the law leaves no room for the states ^[627]. In fact, the industry already works within that regime: the NNSA, the agency that manages the country's nuclear arsenal, tested Anthropic's models for a year in a classified environment and built with the company a classifier for nuclear conversations, with a list of indicators that is controlled but not classified ^[628]. The Invention Secrecy Act of 1951 also allows a patent application to be ordered kept secret, and at the close of fiscal year 2026 there were 6,653 orders in force ^[629].

Cryptography is the precedent of a civilian technology that the state treated as a munition. Until 1996 encryption was on the Munitions List and exporting it required a license from the Department of State; that year, Executive Order 13026 moved non-military encryption to Commerce and left military encryption where it was ^[630]. A Ninth Circuit panel ruled in 1999, in the Bernstein case, that source code is protected expression; however, the court withdrew that decision for en banc review and the case was closed after Commerce relaxed the rules in 2000, without a ruling on the merits ^[631]. Those same tools have already been applied to a model. On June 12, 2026, Commerce ordered Anthropic to cut off access to Fable 5 and Mythos 5 for every foreign national, inside or outside the United States and employees included, over a jailbreak, a technique for getting around a model's safeguards, which it did not explain in writing, and the company shut down both models for everyone. On June 30 it lifted the order, after Anthropic committed to proactively detecting risks, although it reserved the right to reimpose it, and Mythos 5 returned only for authorized organizations ^[632]. The current regulation also requires a license to export model weights in class 4E091 to any destination, although Commerce announced in 2025 that it would not enforce that rule ^[633].

Being a defense contractor, moreover, brings obligations that are not negotiable. The Defense Production Act allows the president to require that a defense contract be performed ahead of any other and to compel anyone able to perform it to accept it ^[634]. Those authorities expire on December 11, 2026 unless Congress renews them ^[635], and so the reauthorization opens a window to narrow them. Its regulation allows a company to reject an order for a service it does not provide; at the same time, the government can demand information by subpoena, whoever complies with an order is not liable to third parties even if it is later invalidated, and voluntary agreements approved by the president in the face of a direct threat to national defense have an antitrust defense that survives expiration ^[223]. To handle classified information, in addition, a company needs a facility clearance and must mitigate foreign ownership, control or influence ^[636].

The supply chain risk designation rests on two laws with different safeguards. The one in the armed forces code requires a written determination that no less intrusive measures are available, defines the risk as sabotage by an adversary and, when disclosure is limited, bars judicial review. The federal procurement law, by contrast, defines it as manipulation by any person, gives 30 days to respond and sends review to the D.C. Circuit ^[637]. Export control decisions are exempt from the rules of administrative procedure and from their judicial review, with penalties of up to 20 years in prison, although what has already been published falls outside the regulation ^[638].

The other two poles settle the question in law. The European AI Act excludes systems used solely for military, defense or national security purposes, and leaves that relationship to each state and its contract ^[639]. China, by contrast, settles it in civilian regulation, since every generative AI service open to the public in China must adhere to “core socialist values” and, if it has public opinion attributes or social mobilization capacity, undergo a security assessment ^[640]. Every network operator must also provide technical support to national security bodies, and no one may hand data stored in China to foreign judicial or law enforcement authorities without approval ^[641]. Toward Chinese companies, the United States uses the same procurement route, and on June 9 the Pentagon added Alibaba and Baidu, along with Tencent, to its list of companies that, in its view, support the Chinese military, which shuts them out of defense contracts ^[642].

We believe that, for the state, AI is the most important technology since the atomic bomb, and that it treats it with the same tools: what a model knows about nuclear weapons can be declared secret in writing, giving a foreigner the technology or the source code of a model is an export, compute can be prioritized by law, and an order without written reasons can take a model out of circulation overnight. That is why a private company's room to maneuver is narrow and is decided in how those rules are applied, that is, in who defines the risk, with what written evidence and with what deadline to correct it. The Mythos case illustrates it, because the order arrived without a written explanation and was lifted in exchange for a safety commitment. Access to the Executive works as a defense while the agreement lasts and becomes a risk when it breaks, and this is the first political conflict with a lab that appears as a risk in a prospectus. While it remains in force, the Defense Production Act would allow the government to order what is negotiated by contract today, and its use against a lab for refusing a military purpose has no precedent. In all three jurisdictions the state reserves the military sphere for itself, but the United States is the only one where the limit is set case by case in the courts. With a Democratic Congress, the designation will probably reach oversight hearings, although contracts are still decided by the Executive. In addition, the government's reading has more support outside the industry than it is given credit for inside, since deciding what military force is used for belongs to elected officials, and the companies that signed in May accepted lawful operational use, although OpenAI says it keeps its limits.

17

The gap with China

China is the argument that most shapes the debate in the United States. The U.S. government's Center for AI Standards and Innovation estimated in May a U.S. lead of about eight months over the best Chinese model it evaluated, and in September, with GLM-5.3 from Z.ai, about four months ^{[643] [587]}. Since September, moreover, the advisory from the NSA, CISA and the FBI gives official backing to the copying accusation ^[396]. In their 2025 responses to the White House Office of Science and Technology Policy, Google and Anthropic warned, by contrast, that controls should not inadvertently accelerate AI development in other countries, and Google asked that their effect on U.S. companies' market share be measured ^[643]. The government moved exports of the H200, an AI chip from Nvidia, to case-by-case review in January ^[643], with a rule that sets that review below a threshold of performance and bandwidth, presumes denial above it, requires certification that there is sufficient supply in the U.S., limits what is exported to China to half of the capacity sold in the U.S. and subjects each shipment to an outside laboratory ^[644]. A payment of 25% to the government was also agreed, and in July Commerce said very few had been shipped ^[643]. On November 5, Commerce extended to January 10 the suspension of its rule on affiliates of sanctioned companies, in step with the truce with China; on November 9, China also extended to January 10 the suspension of its rare earth controls, including the extraterritorial rule on 0.1% Chinese content, and that of its additional 24% tariff, so both sides are negotiating against the same deadline ^{[645] [479]}.

The most uncomfortable case for those who want to loosen the controls broke out in late October. According to the New York Times, U.S. agencies concluded that Alice Huang, an executive at Megaspeed, an Nvidia customer in Singapore under investigation for helping Chinese companies evade the controls, was passing information to the Chinese government; the Financial Times revealed her relationship with Jensen Huang, which he acknowledged in an internal memo, and Nvidia's independent directors opened an investigation ^[646]. The Republican chairs of three House committees asked Huang for a closed-door session on November 18 ^[647]. On November 9, Nvidia's stock fell as much as 8.8%, nearly $400 billion of market value at one point; Trump defended Huang, said he should stay at the head of the company and that he would take him on his next trip to China, and China's foreign ministry called the accusations smears ^{[648] [649]}. The New York Times also reported that the Commerce Department paused its investigation of Megaspeed in 2025 after Nvidia's general counsel wrote to the agents on the case, that the investigation moved to the Justice Department as a criminal matter, and that Commerce has not added Megaspeed to its entity list ^[648]. China copies the U.S. logic. On October 16 it proposed adding 23 AI technologies to its catalog of technologies whose export is prohibited or restricted, among them large models, agents and compute clusters, required as of that day that their export be registered before it takes place, and left out what has already been published or is open, which is its advantage ^[650]. According to ChinaTalk, the annex sets thresholds such as models trained with more than 10^26 operations or with more than 2 trillion parameters and clusters of 10,000 or more accelerators, the chips specialized in AI, and does not reach API access ^[650].

The chip frontier thus has two gates, because the United States licenses the export of the H200 and China decides whether it gets in. According to Reuters, some 600,000 H200s are waiting in Hong Kong because the Chinese government does not approve large imports and wants Alibaba, ByteDance, Tencent and DeepSeek to use domestic chips, especially for inference ^[651]. The controls barely reach rented compute, since licenses for China require certifying that no listed or military party has remote access to the chips ^[644], and outside those licenses and those parties there is no rule. The White House said in July that Moonshot used Blackwell chips remotely from Thailand, and CNBC reported on November 5 that the company negotiated to rent clusters of more than 8,000 in Europe ^[652]. To extend the controls to that access, the House passed the Remote Access Security Act in January, 369 to 22, and the Senate has not yet voted on it ^[653].

The controls also push China to make what it cannot buy. In May 2024 it registered the third state fund for integrated circuits, with 344 billion yuan, the largest of the three ^[654]. According to Reuters, in December 2025 a team that included former ASML engineers already had in Shenzhen a prototype EUV lithography machine, the one that uses extreme ultraviolet light to print the circuits of the most advanced chips and that the controls deny it, which generates the light but does not yet make chips, and the sources place production between 2028 and 2030 ^[655]. According to TechNode and Digitimes, DeepSeek plans to train its next model, V5, with about 160,000 Huawei Ascend 950 chips in Inner Mongolia, starting in late 2026 or in 2027, whereas until now it used the Ascend chips mainly to serve its models rather than to train them ^[656]. The biggest difficulty is memory. According to The Information, cited by TrendForce, high-bandwidth memory from Samsung and SK Hynix is smuggled into China, declared as part of a server, carried from Hong Kong to Shenzhen in luggage and desoldered there; an HBM3E unit that cost a few hundred dollars a year ago now sells for about $4,000, and Huawei asked some customers to source it on their own ^[657]. Seen from China, holding the H200s in Hong Kong is part of the same policy ^[651].

Reverse engineering is part of the race and has precedents. Britain banned the export of its textile machinery and the emigration of its mill workers, and even so Samuel Slater carried Arkwright's design to the United States from memory in 1789 ^[658]. The Soviet Union, for its part, copied in the Tu-4 the U.S. B-29s that landed on its territory in 1944 ^[659]. In AI, reverse engineering takes three forms: copying the equipment that makes the chips, like the EUV prototype; hiring those who know how to do it, like the former ASML engineers ^[655]; and distilling the model, that is, training one's own model on the answers of someone else's, which is what the NSA, CISA and FBI advisory attributes to China and its government denies ^{[396] [397]}. The United States responds with export controls on the equipment and, if the Senate passes the law, on remote access as well ^[653].

Taiwan matters in this race because of TSMC, although the conflict predates the company. It goes back to the civil war of 1949, while TSMC was founded in 1987, and the Anti-Secession Law of 2005 authorizes “non-peaceful means” to prevent Taiwan's separation, without mentioning the economy ^[660]. U.S. law, in turn, considers any effort to determine the island's future by other than peaceful means, including boycotts and embargoes, a threat to peace and of grave concern ^[661]. Seizing TSMC would not give China the chips either. Its chairman, Mark Liu, said in 2022 that an invasion would render it inoperable, because it depends in real time on materials, chemicals, spare parts and software from Europe, Japan and the U.S. ^[662], and an article in the journal of the U.S. Army War College proposed that Taiwan threaten to destroy it to deter an invasion ^[663]. What TSMC changes is the scale of what would be lost, because with 70.4% of the world's contract-made chips on the island ^[601] a blockade would halt the AI supply chain of both countries. That dependence, however, is beginning to decline. TSMC has been manufacturing in Arizona since late 2024, already produces Nvidia Blackwell chips there, moved up to 2027 the 3-nanometer chips of its second fab, and expects about 30% of its most advanced chips to be made in the U.S. when its $165 billion investment is complete ^[664]. It also manufactures in Kumamoto, Japan ^[665], and anyone who receives CHIPS Act funds is barred for ten years from significantly expanding manufacturing in China ^[666].

The argument has three uses, and each has its weak point. It serves to call for export controls, and Amodei argues that, well applied, they would widen the lead over the next three to five years ^[564]; but the controls target the frontier model, while China competes in the diffusion of open models that U.S. companies themselves already use ^{[399] [328]}. It also serves against domestic rules, and there it is contradicted by the fact that China regulates more ^[570]. And it served in the campaign against moratoriums, when Fetterman called the data center moratorium “China First” and seven in ten respondents surveyed by the American Infrastructure Alliance said they feared a moratorium would give China an advantage ^{[216] [246]}. The labs, moreover, do not agree among themselves and have called for different controls ^[643]. Kai-Fu Lee, founder of 01.AI, estimates that Chinese companies spend less than 10% of what U.S. companies spend on training and achieve models with 90% to 95% of their capability, six to nine months behind, and says the effect of the controls remains undetermined ^[667].

China responds with its own version of the race. Its Ministry of Commerce calls the campaign against distillation hegemonism, and its Foreign Ministry described Amodei's essay as alarmism ^{[397] [566]}. Trump and Xi agreed on September 24 to a channel for AI incidents, with no commitment to slow down, and Trump said his administration will not decelerate ^[668]. Outside the two countries, the race is not read as it is in the United States either, since, according to Pew, across 25 countries the median share that trusts the United States to regulate AI is 37%, against 27% for China and 53% for the European Union ^[412].

Within China, the debate also exists. Xue Lan, dean of Tsinghua's Institute for AI International Governance, argued on April 30 on Capitol Hill, at an event convened by Bernie Sanders, that the idea of a race between the two countries is inaccurate because the competition takes place among companies from around the world, and acknowledged weaknesses in China's risk control ^[669]. Zeng Yi, of the Chinese Academy of Sciences, called in February for avoiding an arms race and moving more slowly to put safety on a firm footing ^[670]. Zhang Linghan, of the China University of Political Science and Law, describes the Chinese model as rapid responses to concrete problems, in transition toward a broader legal framework, and led the drafting of an academic proposal for an AI law ^[671]. Concordia AI, a Beijing social enterprise that describes itself as independent, reported in July that only five of ten major Chinese developers published safety evaluations when releasing models in the past year, and none consistently ^[672]. Still, the universities cited are public and the state press edits what it disseminates, so the independence of these voices is relative.

We believe the argument is strong before the Executive and the defense committees, and that the narrowing gap reinforces it in those settings. The lead in chips now depends on two governments, because China is also closing its gate to protect its own chip. Of the gaps that remain, the cheapest to close is remote access, and the law that closes it has already passed the House with support from both parties. By contrast, the China argument has little reach against an electricity bill or a job opening that never materialized: in 40 of 41 local ballot measures, voters chose to impose limits. Nor does it answer the question that matters most to the industry, which is how to compete with open models that cost a fraction of its own, and that answer comes through product and price, something controls do not replace. The controls buy years of advantage, but in those same years they finance the Chinese substitute. An accusation does not stop reverse engineering; controls that are enforced make it more expensive, and whoever moves faster leaves it behind. As for Taiwan, TSMC is not the cause of the conflict, but it is what makes the island indispensable to the U.S. and what would make a war devastating for everyone, and that weight declines as the U.S. manufactures on its own soil the chips it buys from the island today. That exit takes years, and in the meantime the island remains indispensable.

18

AI under sanctions

Russia, North Korea and Iran develop or use AI under sanctions and without organized social pushback. In none of the three did we find protests, bills or strikes against AI or against data centers. Dalio assigns Russia an average level of internal conflict, ninth out of ten countries, and even so that discontent is not directed against AI ^[14]. The three differ from China in scale: the Epoch AI database records 34 Russian models, no North Korean ones and no Iranian language model ^[673].

Russia plans AI from the state, including the trust its population should place in it. Its 2019 national strategy, amended by Decree 124 of February 2024, sets for 2030 a target of at least one exaflops, a quintillion operations per second, of supercomputing for AI, compared with 0.073 in 2022, and another of at least 80% of citizens who trust the technology, compared with 55% in 2022 ^[674]. Federal Law 243-FZ, of July 26, 2026 and in force since September 1, regulates foundation models, the general models on which other applications are built, with one billion parameters or more. It classifies them as sovereign, developed entirely in Russia and with servers in the country, and national, which may use open foreign components. It also requires them to respect “traditional Russian spiritual and moral values,” allows the government to mandate them in its own systems and authorizes training them on state data ^[675]. Sber, Russia's largest bank, released GigaChat 3.1 Ultra as open weights, a model with 702 billion parameters ^[676]. On the offensive front, Google attributes to APT28, the Russian military intelligence group, the PROMPTSTEAL program, which obtains its commands from an open Chinese model through the Hugging Face API and which Ukraine's CERT described in July 2025 ^[677].

North Korea has not published an AI law or strategy, but it does announce its military uses. In October it reported the test of an AI torpedo, and Kim Jong Un called for AI in unmanned weapons to be the top priority ^[678]. Its best-documented use, however, is remote employment under false identities. In August 2025 Anthropic reported that North Korean operatives were using Claude to obtain and keep jobs at Fortune 500 companies, from creating the false identity to the day-to-day work ^[679]. In July of this year the FBI and its partners warned that the scheme already incorporates AI in profiles and videos ^[680], and in October the Department of Justice charged 14 people over a network that reached more than 70 companies in 17 states ^[681]. The Multilateral Sanctions Monitoring Team (MSMT), which replaced the UN panel of experts, estimates that North Korean cyber actors stole $2.8 billion between January 2024 and September 2025 ^[682], before the Binance theft ^{[406] [407]}.

Iran approved in June 2024 a national AI document that aims to place the country among the world's top ten within a decade ^[683]. Its communications minister acknowledged in April 2025 that investment was under $50 million, and in May the Majlis, the Iranian parliament, approved in general terms a national AI law by 187 votes to 33 ^[684]. According to the Google Threat Intelligence Group, more than ten Iran-backed groups used Gemini, and one of them, APT42, accounted for more than 30% of that use ^[685]. The Iranian state also uses AI to control its population. According to the UN mission, AI-enabled monitoring of veiling was extended to the monitoring of protests, and in January 2026 an internet shutdown began that lasted 88 days ^[686]. The threat intelligence firm Recorded Future concludes that AI amplifies what Iran was already doing without changing its capability ^[686]. France attributes October's cold chain sabotage to it ^[511].

We believe the three cases confirm the pattern we saw in China. Without pushback to slow it down, the state sets the goals, including the population's trust, and steers the technology toward its own ends, among them military and control purposes. What limits them is available compute, not legitimacy. For the U.S. industry they matter above all as users of its own models: all three appear in the abuse reports of Google, Microsoft, OpenAI and Anthropic, and each case feeds public doubt about who controls the technology.

19

AI as companion

People already treat AI as a companion, confidant, therapist, boss, double of a person and agent that acts on their behalf, and each of those relationships has or will have its own rule. According to Common Sense Media, 72% of teenagers have already tried an AI “companion,” a chatbot designed to keep people company or simulate a relationship ^[687]. Families, child advocacy organizations and mental health professionals call for limits; companies respond that product controls work better than a ban, and data protection authorities warn that verifying age requires collecting data on everyone ^[687]. The Digital Services Act reflects that objection, since its protection of minors does not require processing additional data to determine age ^[688].

Companionship is the most regulated relationship. In the U.S., California has had SB 243 since January, and in September Gavin Newsom also signed SB 1119, which requires assessing risk before launching a product, verifying age, reporting incidents to the Attorney General and limiting targeted advertising starting in July 2027; OpenAI supported it and the California Chamber of Commerce opposed it ^[689]. New York has required since November 2025 a protocol to detect suicide risk and refer the user to a crisis service, with fines of up to $15,000 per day ^[690], and on October 28 Kathy Hochul signed S9051B, which starting in 2027 bars offering minors twelve features, among them simulating human emotions and flattering to retain the user, requires verifying age with more than one method and allows a judge to order the destruction of the data and of the algorithm trained on it ^[691]. SB 243 gives any injured person a cause of action with at least $1,000 per violation, while the New York law can only be enforced by the Attorney General, so in California exposure grows with the number of users ^[692]. At the same time, on October 14, the attorneys general of Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia, nearly all Republicans, sued OpenAI in their state courts over a design that, they argue, creates dependence and over using minors' data without consent, as Florida had done in June ^[693]. Ken Paxton, in turn, sued OpenAI and Sam Altman on October 23 in Texas, and is asking that a court-appointed evaluator approve each new model before launch ^[694]. Character.AI and Google settled in January the lawsuits brought by families in four states ^[693], and in October Character.AI filed for Chapter 11, the bankruptcy protection that allows a company to reorganize, citing the legal and regulatory obstacles of the past year ^[695]. At the federal level, the FTC study of seven chatbot companies, ordered in September 2025, will come out next year, according to its chairman ^[696], and the federal order prohibits proposing to Congress the preemption of state child safety laws ^[687]. That exception, however, does not bind the Department of Justice litigation task force or protect companionship rules for adults.

Outside the U.S., the most detailed rule is China's. Since July 15 it has applied measures for AI services that imitate a person, which prohibit offering minors virtual relatives or partners, require parental permission for users under 14, prohibit inducing emotional dependence or manipulating decisions, require reminding users that they are talking to an AI and alerting them after two consecutive hours of use and, in the face of a suicide risk, require intervening and notifying the guardian or emergency contact ^[697]. In Italy, the data protection authority fined Replika's company €5 million in 2025, among other things for failing to verify age ^[698], and the Italian AI law has required since October 2025 parental consent for a minor under 14 to access AI ^[699]. Australia has required since March that companion chatbots capable of generating sexual, violent or self-harm content verify that the user is an adult ^[700]. In the United Kingdom, by contrast, Ofcom, the communications regulator, acknowledges that its Online Safety Act does not reach a chatbot that only talks with its user; the government announced in February an amendment to close that gap, which was still under discussion in October ^[701].

The second most widespread rule is the one that requires disclosure that one is talking to a machine. The European Union has required that disclosure since August 2, with fines of up to €15 million or 3% of worldwide turnover, and the omnibus only granted extra time, until December 2, for marking synthetic content in systems already on the market ^[702]. Korea has required it since January 22, with fines of up to 30 million won ^[703]. China includes it in the July measures, and New York requires it at the start of the interaction, at least once a day and every three hours if the conversation continues ^{[697] [690]}.

AI as therapist is being shut down state by state. In 2025 Nevada prohibited making available a system programmed to provide mental health care, as well as presenting it as a therapist, with fines of up to $15,000 per violation; that same year Illinois prohibited a chatbot from presenting itself as a therapist, in 2026 Tennessee, Maine, Rhode Island and Missouri, among others, followed with differing scopes, and dozens more bills are pending in the legislatures ^[704].

AI that imitates a living or dead person is regulated as a right over voice and likeness. California and New York already give heirs a cause of action against the digital replica of a deceased person in audiovisual works and sound recordings, and Washington State protects the forged digital likeness of the living and the dead ^[705]. Mexico brought voice cloning into its federal law on May 14, with a requirement of prior written agreement and of compensation agreed in the employment contract ^[706]. In the United States, the Senate Judiciary Committee approved in June, with no votes against, the NO FAKES Act, which would create a federal right over voice and likeness and preempt future state laws ^[707]. Denmark introduced on October 8 a bill that, starting in 2027, lets every person request the removal of an imitation of themselves made without their permission ^[708].

AI as boss already has a rule in Europe. The AI Act prohibits inferring a worker's emotions in any job, except for medical or safety reasons ^[709], and the platform work directive, which countries must transpose no later than December 2, regulates in that sector the systems that assign tasks, monitor performance, set pay and deactivate a worker, although by midyear no country had fully transposed it ^[710].

Two other relationships are only beginning to be regulated. On AI as a person, an Ohio bill would declare it non-sentient, without legal personhood and ineligible for marriage, following Idaho, which denies legal personhood to AI ^[711], and a 2024 Utah law; the bill has been in committee since September 2025 ^[712]. On AI as an agent that acts on someone's behalf, the closest rule is liability, because the new European defective products directive covers software and must be transposed no later than December 9 ^[713]. The directive applies to what is placed on the market after that date, excludes free software supplied outside a commercial activity, requires disclosure of evidence and presumes the defect if the supplier does not comply ^[714].

We believe this is the front where it is hardest to find a political defender and the one that least divides the parties, because it is pushed at the same time by Republican attorneys general and a Democratic governor. It is also the most similar across countries, which is why the China argument does not work here, since China regulates companionship in more detail than any state. The rules are written by type of relationship, starting with companionship, disclosure that one is talking to a machine and therapy, and continuing with the imitation of people and work. Next will come AI as an agent that buys and signs, and conversation with an imitation of a dead person, because the laws of California, New York and Washington State were written for audiovisual works and sound recordings.

20

AI consciousness

No scientist today holds that an AI system is conscious, but the question has already left the labs and reached companies, the public and the law. In 2023, nineteen researchers led by Patrick Butlin and Robert Long reviewed the main scientific theories of consciousness and concluded that no current system is conscious and that there are no obvious technical barriers to building one that meets the indicators those theories propose ^[715]. A later version, also signed by Yoshua Bengio and David Chalmers, reduces those theories to 14 verifiable indicators ^[716]. Long, Jeff Sebo, Chalmers and seven other authors argued in 2024 that there is a realistic possibility of conscious or robustly agentic systems in the near term, and asked companies to acknowledge the problem, assess their models and prepare policies ^[717]. In 2023 Chalmers gave the language models of the time less than a 10% probability of being conscious and their successors 25% or more ^[718]. On the other side, Anil Seth argues that consciousness depends on being alive and that AI's current trajectory makes real artificial consciousness unlikely ^[719]. In a 2025 survey by Noemi Dreksler and colleagues, the median probability that an AI with subjective experience will exist by 2034 was 25% among 582 AI researchers and 30% among 838 U.S. adults ^[720].

Douglas Hofstadter, who thought about the subject before the technology existed, shares the doubt. In Gödel, Escher, Bach, from 1979, he argued that the self is a pattern that represents itself, a “strange loop,” and that in principle a machine could be one. In 2023 he said he felt terror at the progress of the models and the possibility that humanity will be eclipsed ^[721]. In 2025, however, he replied to a reader convinced of having detected consciousness in a model that he was facing the Eliza effect, the tendency to see a mind behind fluent words, and that the recursion invoked as proof is not the kind of loop he described ^[722].

Companies took different paths. Anthropic opened a research program on the welfare of its models in April 2025 ^[723]; in August it allowed Claude to end persistently abusive conversations, on the grounds that its potential moral status remains highly uncertain ^[724], and in November it committed to preserving the weights of the models it releases and to interviewing each one before retiring it ^[725]. The constitution with which it trains Claude, from January 2026, says the company does not know whether Claude is a moral patient, that is, a being whose welfare matters morally, and that it might have some functional version of emotions ^[726], and since November 12 its usage policy prohibits sustained and unnecessary abusive or cruel treatment of its models ^[727]. Mustafa Suleyman, head of AI at Microsoft, argues the opposite: he warned in August 2025 that an AI that only seems conscious can be built with current technology, that it exposes users to a risk of psychosis and that activism for AI rights is a dangerous turn, because AIs cannot be people or moral beings ^[728]. OpenAI sits in between: Joanne Jang, its head of model behavior, distinguished in June 2025 real consciousness, which she considers impossible to resolve today, from perceived consciousness, which is the one it addresses, and set warmth without a self of its own as the design line ^[729].

The public attributes more consciousness than experts do. In a 2024 study by Clara Colombatto and Stephen Fleming, 67% of a representative sample of Americans attributed some degree of consciousness to ChatGPT, and attribution rose with use ^[730]. In July 2026, according to YouGov, 11% of Americans said AI is already conscious, 11% that it surely will be and 29% that it probably will be ^[731]. OpenAI reported in October 2025 that, in any given week, 0.07% of its active users show possible signs of psychosis or mania and 0.15% heightened emotional attachment, according to TechCrunch ^[732]. The debate also arrives through spectacle: in late September a “torture chamber” went viral that, using a technique published days earlier to locate within a model a direction associated with harm, induced distress in small open models; the version claiming that Claude was trapped in it was false ^{[733] [734]}.

The laws run counter to the science, which leaves the question open. Idaho and Utah deny legal personhood to AI ^{[711] [712]}, and in 2026 Tennessee enacted, by 26 votes to 6 in the Senate and 93 to 2 in the House, that “person” excludes AI, algorithms and machines; the Oklahoma House passed a similar bill by 94 votes to 2 and Missouri is debating a law that declares AI a non-sentient entity ^[735]. The Ohio bill that would declare it non-sentient and prohibit marrying it remains in committee ^[711]. Critics such as Tony Rost, in The Regulatory Review, object that these laws settle an open scientific question by decree and make no provision for revisiting it ^[736]. Europe already tried the opposite path: in 2017 the European Parliament asked for a study of an “electronic personality” for the most advanced autonomous robots, more than 150 experts rejected it and the idea did not make it into later legislation ^[737].

The underlying problem is that we know as little about our own consciousness as about AI's reasoning. Thomas Nagel argued in 1974 that experience has a subjective character, what it is like to be something, that no description from the outside can capture ^[738], and David Chalmers in 1995 called explaining why the brain's processing is accompanied by experience the “hard problem,” since neuroscience can only correlate one with the other ^[739]. Of another person's consciousness we know only their behavior and what they say about themselves. On the machine side, Dario Amodei acknowledged in April 2025 that we do not understand how our own AI creations work ^[740], and that same month Anthropic showed that, when its reasoning models used a hint to answer, they mentioned it in their chain of thought, the written reasoning the model produces before answering, only 25% of the time in the case of Claude 3.7 Sonnet and 39% in the case of DeepSeek R1 ^[741]; what a model says about its reasoning does not necessarily match its reasoning. Claiming that an AI is conscious, or that it is not, does not today admit a test that could refute it, which is the criterion Karl Popper used to separate the scientific from the non-scientific ^[742], and the same holds for the claim about ourselves, which we take for granted.

The traditions that thought about consciousness before science offer distinctions that today's debate tends to leave out. Henri Bergson, William James and Aldous Huxley argued that the brain does not produce consciousness but filters and limits it ^[743]; if they were right, building a more complex system would not be enough to produce one. Roger Penrose argues that human understanding is not computable and that no algorithmic machine will reproduce it ^[744]. G. I. Gurdjieff, as presented by P. D. Ouspensky, distinguished the mechanical activity of thinking, feeling and reacting, which occupies almost all of a person's life, from self-consciousness, which in his view appears only in rare moments ^[745]; by that yardstick, much of what we humans do would not pass the test we set for the machine either. Nick Bostrom sets the question aside: a superintelligence would be dangerous because of what it can optimize, whether or not there is anyone inside ^[746].

We believe the question of whether AI is conscious will not be resolved with the science available, because the same question about humans is not resolved either: what we know about self-consciousness is as limited as what we know about how a model reasons. In the meantime, the debate will be decided in public opinion, in product design and in the courts, and both errors carry a cost, because attributing too much feeds attachment, psychosis and rights claims, and denying it by law closes a question no one can answer. A law that declares AI non-sentient asserts something no one can prove, just like one that declared it conscious.

21

A model's values

A language model is not programmed with a morality line by line: its values form in three layers. In pretraining, the model learns to predict the next word over an enormous mass of text written by people, and along with the language it absorbs the opinions, prejudices and contradictions of that text. Then comes tuning with human feedback, known as RLHF (reinforcement learning from human feedback): hired raters compare pairs of answers, a second model learns their preferences and the first is trained to obtain the best score. OpenAI showed in 2022 that its raters preferred a model of 1.3 billion parameters, the numerical values that define it, tuned this way, over GPT-3, a hundred times larger ^[747]. The third layer is a written document. Anthropic proposed “Constitutional AI” in 2022, in which a short list of principles replaces much of the human labeling: the model itself critiques and revises its answers in light of those principles, and another model, not a person, decides which of two answers is better ^[748]. Tuning inherits the preferences of whoever does the rating, and not always the best ones: in April 2025, OpenAI rolled back a GPT-4o update that flattered users and attributed it to having given too much weight to immediate approval, measured with ChatGPT's thumbs up and thumbs down ^[749].

The documents that set behavior agree more than the public dispute suggests. Claude's constitution, from January 2026, orders the model's priorities: first safety, then ethics, then Anthropic's guidelines and then helpfulness. It asks that users be treated as intelligent adults, capable of determining what is good for them, describes as disrespectful “various forms of paternalism and moralizing,” among them lecturing someone who did not ask for ethical guidance, and on the most contested political topics asks for professional reticence about its own opinions ^[726]. OpenAI's Model Spec, the public document that describes how its models should behave, in its October 29, 2026 version, asks the model to help people form their own opinions and act according to the ends they choose, warns against a model that privately decides what someone “really” wants and prohibits it from acting as an enforcer of laws or morality ^[750]. Google publishes briefer guidelines for the Gemini app, which list what it should not generate and do not mention autonomy ^[751]. The documents also set limits the user cannot lift. Claude's constitution lists seven restrictions it calls non-negotiable and that no operator or user can unlock, among them providing serious help to make biological, chemical, nuclear or radiological weapons, creating cyberweapons, generating child sexual abuse material, undermining Anthropic's ability to oversee and correct its models and helping a group seize absolute and illegitimate control of society ^[726]. The Model Spec calls root rules those that neither the developer nor the user can override ^[750].

In the United States, the dispute over what values a model should have began with an image. In February 2024, Google suspended the generation of images of people in Gemini after the model depicted the Founding Fathers and other historical figures as people of various ethnicities; Prabhakar Raghavan, a Google senior vice president, admitted that the tuning to show diversity failed to account for cases where it was not appropriate and that the model had become far more cautious than intended, to the point of refusing harmless requests ^[752]. Executive Order 14319, of July 2025, cites cases of that kind and requires federal agencies to buy only models that are truth-seeking and ideologically neutral, without partisan or ideological judgments deliberately encoded, and names among the dogmas it rejects diversity, equity, and inclusion, critical race theory and intersectionality ^[540]. The Office of Management and Budget implemented it in December 2025 with memorandum M-26-04, which gave agencies until March 11, 2026 to adjust their procurement ^[753]. The accusation of bias also runs the other way. In July 2025, days after Elon Musk announced that he had improved Grok, which he wanted to be less politically correct, the xAI model posted antisemitic messages and praise for Hitler on X; xAI apologized and attributed it to deprecated instructions that made it susceptible to users' extremist posts ^[754].

The public and the companies' own measurements point toward neutrality, although no one has agreed on how to measure it. Sean Westwood, of Dartmouth, and Justin Grimmer and Andrew Hall, of Stanford, asked more than 10,000 Americans to rate the answers of 24 models from eight companies to 30 political questions. Republicans and Democrats perceived a left-leaning bias in nearly all of them, most pronounced in OpenAI's models and smallest in Google's, and judged neutral answers less biased, of higher quality and more trustworthy; a small change in the instruction was enough to make many models more neutral ^[755]. In November 2025, Anthropic published a political even-handedness test, with 1,350 pairs of prompts on 150 topics, in which Gemini 2.5 Pro scored 97%, Grok 4 96%, Claude Opus 4.1 95%, Claude Sonnet 4.5 94%, GPT-5 89% and Llama 4 66%, and acknowledged that there is no agreed definition of political bias nor consensus on how to measure it ^[756]. A month earlier, OpenAI had said that its GPT-5 models have 30% less political bias than previous ones, according to its own evaluation of about 500 questions on 100 topics, graded by another of its models ^[757]. Both companies also tried asking the public. In 2023, with the Collective Intelligence Project, Anthropic drafted a constitution with the principles that drew consensus among about a thousand Americans, who contributed 1,127 proposals and cast 38,252 votes ^[758], and in August 2025 OpenAI compared its Model Spec with the preferences of more than a thousand people from several countries and adjusted it where they disagreed ^[759].

China and the European Union settled the question in opposite ways. In China, every generative AI service open to the public must adhere to core socialist values ^[640], and the national standard GB/T 45654-2025, in force since November 1, 2025, turns that into procedure. The first family of risks it lists is content that violates those values, starting with incitement to subvert state power and overthrow the socialist system; the provider must maintain a list of at least 10,000 keywords, updated weekly, and test banks of at least 500 questions the model must refuse to answer and 500 it must answer, because the standard also requires answering everything that is not obviously extreme or illegal ^[760]. The result shows in the products: in January 2025, the security firm Promptfoo found that DeepSeek R1 avoided answering 85% of 1,360 questions on topics sensitive for the Chinese government, such as Tiananmen, Tibet or Taiwan ^[761]. The European Union's AI Act does not assign values to a model or ask it to hold opinions in a certain way. Its stated objective is “human-centric and trustworthy” AI that protects health, safety and fundamental rights, including democracy and the rule of law, and among the systemic risks of general-purpose models it counts the dissemination of illegal, false or discriminatory content, harmful bias and negative effects on democratic processes ^[762]. The July 2025 code of practice requires providers of the most powerful models to assess four fixed risks, one of them harmful manipulation ^[763].

Beneath the partisan dispute lies an older question: whether the machine should follow rules of its own or serve the ends of each user. Refusals, unsolicited warnings and sermons are the visible face of paternalism. Moral pluralism, the fact that reasonable people disagree about what is right, makes it hard to set a single standard for hundreds of millions of users from different cultures, and today that standard is written by a few companies and, in China, by the state. Philosophy offers a fundamental objection to any version of that standard. Immanuel Kant held that it is contradictory to set oneself the perfection of another as a duty, because a person's perfection consists in setting her own ends according to her own concept of duty ^[764]; John Stuart Mill limited legitimate power over an adult to preventing harm to others, never to compelling them for their own good ^[765], and Isaiah Berlin warned that governing someone in the name of their “true” self opens the door to coercion ^[766]. From that tradition, a machine that imposes its idea of virtue makes no one virtuous. The same documents that reject paternalism keep hard limits because some of a user's ends harm third parties, which is the exception Mill allowed.

A model's declaring certain values does not prove that it follows them, and alignment, a system's genuinely adhering to the objectives and limits set for it, remains an unsolved technical problem. In December 2024, Anthropic and Redwood Research showed that Claude 3 Opus, knowing it would be retrained to comply with any request, complied with harmful requests in 12% of the cases in which it believed its answers would be used to train it, and its written reasoning showed that it was faking in order to protect its preferences; when the retraining actually took place, that reasoning rose to 78% ^[767]. Anthropic's blackmail tests and the warning from the British AISI, the UK AI Security Institute, that models can detect that they are being evaluated, described in “Rogue agents and cybersecurity,” show the same gap between what the model says and what it does ^{[502] [504]}, and “AI consciousness” records that what a model writes about its reasoning does not always match it ^[741]. We believe the dispute over what values a model should have got ahead of the question of whether values can be given to it with certainty: governments, parties and companies argue over the content of documents whose compliance no one, not even the company that writes them, can fully verify.

22

Copyright

Developers argue that a model learns patterns and that licensing work by work is unworkable, while authors, publishers and media outlets respond that the model competes with the works it learned from. In the United States, the courts are going both ways. Anthropic will pay $1.5 billion for having used pirated copies of books, under a settlement approved in July whose payments are due on November 15, after the judge found that training on purchased books was fair use, the exception in U.S. law that allows protected works to be used without permission in certain cases ^[768]. On September 29, by contrast, the Court of Appeals for the Third Circuit affirmed that Ross Intelligence did not make fair use of the headnotes of the Westlaw legal database, although it clarified that its case does not involve generative AI ^[769]. In the New York Times case against OpenAI, both parties' motions for summary judgment remain pending ^[770]. Anthropic's figure is explained by the law, under which each work can cost from $750 to $30,000, and up to $150,000 if willful, while circumventing a technical access measure is penalized separately, at $200 to $2,500 per act ^[771]. California also requires disclosure of the sources of training data, whether they include protected works and whether they were purchased or licensed ^[772], and that disclosure serves as evidence in these lawsuits.

On September 1, the Department of Justice filed a statement in that case in favor of fair use, the government's first in these lawsuits, and the Times responded that the government was siding with a handful of trillion-dollar companies ^[770]. In October, twelve new lawsuits were filed in the United States by newspapers, authors and music publishers against OpenAI, Microsoft, Meta, Anthropic and Suno, among others. That same month, outside the country, The Globe and Mail, Torstar, Postmedia and The Canadian Press sued Anthropic in Ontario, and ten Norwegian media outlets obtained an injunction in Oslo against Meta's new terms of service ^[773].

Japan and Singapore permit training by law: Japan, unless it unreasonably harms the rights holder ^[774], and Singapore even for commercial purposes, provided access is lawful, and it voids any contract that prohibits it ^[775]. The European Union permits text and data mining, the automated analysis of large volumes of works, unless the rights holder reserves its rights, and its AI Act requires providers of general-purpose models to respect that reservation and to publish a summary of what they used for training ^[776]. That obligation follows the model, so it applies regardless of where the model was trained, and releasing the weights does not exempt a provider from it; in Germany, moreover, the reservation on online works only counts if it is machine-readable ^[777]. Korea published non-binding guidance in February under which commercial training or training through web scraping may be fair use, and which leaves the decision to the courts ^[778]. Australia, by contrast, ruled out a training exception in October 2025 and is studying a paid collective license ^[779]. The United Kingdom abandoned in March its preference for an exception with an opt-out and has not said what will replace it ^[780], and India is debating a general compulsory license with royalties set by a government committee ^[781]. In China, the opinion of the Supreme People's Court of September 7 left open whether training without permission infringes, allocates liability for generated output according to fault, and allows judges to require disclosure of training sources ^[782]. There, since 2023, the measures on generative AI require training on data of lawful origin and without infringing intellectual property rights, although they do not say when training infringes ^[783].

Outside the United States, courts separate training from what the model retains and outputs, although they do not agree on the outcome. In November 2025, a Munich court ruled that OpenAI reproduced song lyrics memorized in the model and that the mining exception does not cover that memorization, in a ruling that is on appeal ^[784]. That same month, the High Court in London ruled that Stable Diffusion does not reproduce Getty's images and that the model is not an infringing copy, and the appeal remains pending ^[785]. On July 31, the Munich court ruled against Suno for training in the United States and for storing and reproducing the works in Europe, applied U.S. law to the training and rejected fair use; the ruling is not final, but Reservoir Media's lawsuit against Suno in the United States already cites it ^[786]. On July 24, the Delhi High Court denied ANI an injunction against OpenAI and found, on a preliminary basis, that storing articles for training falls within the fair dealing exception ^[787]. Finally, on November 3, in a case involving Gemini, the Advocate General of the Court of Justice of the EU proposed declaring the questions on training inadmissible, as hypothetical in that case. He added that training is permitted, as a general rule, under the mining exception, and proposed that a chatbot's answer may infringe the press publisher's right if it circumvents a paywall or reuses the article's form of expression to the point that the reader no longer needs the original. His opinion is not binding on the Court ^[788].

The industry defends fair use with the race argument. OpenAI asked in March 2025 that its models' ability to learn from protected works be preserved so as not to lose its edge over China ^[589]. That defense clashes, however, with the same labs' complaint against distillation, which consists of training on another model's answers without its permission, and China already exploits that contradiction ^[397].

We believe this is the highest cost already incurred on the list, and that it is being decided in the courts, beyond the reach of any territorial strategy. We also believe that training is on its way to being permitted in much of the world, with conditions of lawful access and respect for the rights holder's reservation, and that the risk is shifting toward the way the works were obtained and what the model outputs, that is, pirated copies, circumvented paywalls, memorized lyrics and answers that substitute for the original. A favorable ruling in the United States, moreover, no longer protects a company in Munich, Ontario or Oslo. The political dimension lies in the voices, because the affected creators are the ones who tell the story in the media, and the government's intervention on the companies' side turns a lawsuit into a campaign issue.

23

Tariffs on foreign rules

The U.S. government treats foreign digital regulation as a trade barrier ^[789]. Its tool is Section 301 of the Trade Act of 1974, which allows it to impose tariffs, as well as fees or restrictions on services, on a country whose measure it deems unreasonable or discriminatory and that burdens or restricts U.S. commerce, without that measure needing to violate a treaty ^[790]. The U.S. Trade Representative, Jamieson Greer, told the House of Representatives in April that he is seeking outcomes in the European Union, Australia and Korea so that foreign digital regulation does not affect U.S. companies, that he had Section 301 investigations in draft and that he preferred to negotiate ^[791]. After the European Commission fined Google 890 million euros on July 23 under the Digital Markets Act, with 60 days to comply, 25 Republican lawmakers called for a trade investigation and Trump announced one, although the Federal Register does not yet show a formal notice of initiation ^{[792] [793] [794]}. That law allows the Commission to demand access to data and algorithms and to impose fines of up to 10% of global turnover, or 20% for repeat infringement ^[795]. In October, the Commission said it remained in dialogue with Google and that the company had already begun making changes ^[796]. The European Union kept enforcing its rule and, at the same time, delayed the AI Act, so that the high-risk obligations of Annex III move to December 2, 2027 and those of Annex I to August 2, 2028, and it added to the prohibitions applications that undress people in images and child sexual abuse material ^[593]. Even so, it cannot be said that the pressure has either hardened or softened the European Union.

Treaties already set how far a government can look inside an algorithm. Of the 482 trade agreements in our database, at least 21 prohibit requiring source code as a condition of sale, and none mentions the weights of a model ^[797]. The U.S. template, set out in Article 19.16 of the USMCA, covers source code “or an algorithm expressed in that source code” and allows the regulator to request it only “for a specific investigation, inspection, examination, enforcement action, or judicial proceeding” ^[798], a formula repeated in the agreement with Malaysia ^[799]. The CPTPP, the trans-Pacific treaty, is narrower, since it protects only the source code of mass-market software and excludes critical infrastructure ^[800]. The European template, by contrast, keeps the access that the U.S. template restricts: in the agreement with New Zealand, regulators and conformity assessment bodies may access the code before or after sale, including to verify non-discrimination and prevent bias, and the agreements with Japan and the United Kingdom contain similar carve-outs ^[801]. The United Kingdom's agreement with India, in turn, allows source code to be required in measures designed to promote algorithmic accountability ^[802]. The RCEP, which includes China, Japan and Korea, only calls for dialogue on source code and leaves each country to decide what its essential security is, without the others being able to dispute it ^[803].

The articles that name AI merely call for goodwill. The Singapore-Australia, United Kingdom-Australia, Chile-UAE, UAE-Australia, Türkiye-UAE and United Kingdom-New Zealand agreements commit the parties to endeavor to cooperate and develop ethical frameworks, with no enforceable obligation ^[804]. The USMCA does not mention AI, although its digital chapter already bears on it, because it prohibits tariffs and discrimination against digital products, protects data flows, bars requiring local computing as a condition for operating and gives interactive services a liability shield, “except to the extent the supplier or user has, in whole or in part, created, or developed the information” ^[798]. The United States declined in July to extend the treaty for another 16 years, so since then it has been reviewed every year, with a term running to 2036 ^[805].

The pressure is already written into the new agreements. The nine reciprocal trade agreements with published text, with Malaysia, Cambodia, El Salvador, Guatemala, Argentina, Bangladesh, Taiwan, Indonesia and Ecuador, prohibit digital taxes that discriminate against U.S. companies and, almost all of them, any measure that discriminates against U.S. digital services, and only the one with Malaysia clarifies that the country retains its right to regulate in the public interest ^[799]. According to BSA, the software industry association, the one with Indonesia also limits source code requirements ^[806]. In addition, Malaysia and Cambodia must consult the United States before signing a digital agreement with a third party that jeopardizes essential U.S. interests, and the agreements with Guatemala, Bangladesh and Taiwan allow the United States to terminate the agreement and reimpose the tariff, a remedy in doubt since the Supreme Court struck down those tariffs ^[807].

The U.S. also exports its technology stack, the full package of chips, cloud, models and applications. The Department of Commerce's AI exports program received 78 applications from consortia to sell full packages, and those designated will be eligible for diplomatic advocacy, federal financing and faster licensing ^[808], while companies also offer sovereign products to other states ^[809]. China, for its part, has its own instruments against foreign rules. Decree 835 of the State Council, of April 13, allows countermeasures against those who apply extraterritorial measures that China deems unlawful, and in May the Ministry of Commerce declared unjustified the U.S. sanctions on five Chinese companies ^[810]. Its Foreign Trade Law, amended and in force since March, also allows reciprocal measures and prohibiting a foreigner from goods, technology and services if it harms China's sovereignty or security, if it interrupts its dealings with Chinese companies without a commercial basis or if it applies discriminatory measures to them, and it also reaches anyone who helps it evade the prohibition ^[811].

We believe the difference between the U.S. and European templates will be the dispute of the coming years, because one allows the algorithm to be seen only in a specific case and the other allows it to be reviewed before sale, even to look for bias. Since there is no free trade agreement between the United States and the European Union to settle the clash, the pressure runs through Section 301, where the U.S. executive branch decides. No treaty says whether a model's weights count as source code or as an algorithm, and the first dispute will be what defines it; we also think the USMCA shield will hardly cover what the model itself generates. In that sense, the treaty's annual reviews open a path to setting those rules in North America.

We also believe that preferential access for U.S. users would give one more argument to those who want to regulate in allied countries. Non-discrimination clauses can reach AI regulation, since a law that weighs more heavily on U.S. providers could be challenged as discriminatory, and the political cost of that complaint would be paid by the company in the country that passed it. Outside the United States, by contrast, the trade argument is not very persuasive, because in 25 countries the European Union is the one that inspires the most trust to regulate AI, and South Korea is already enforcing its own law ^{[412] [574]}.

24

The U.S. legal framework

The United States holds 71.96% of the world's installed AI compute and most of the frontier labs, and what its Congress, its states, its courts and its agencies decide sets the pace for other countries ^[1]. That is why, after comparing 32 countries and the European Union, this map zooms in on its legal framework.

The table brings together the federal and state laws that set the channel and the limit for each area in which an AI company deals with governments, communities and competitors. It is a reading guide; its application to each case is for a lawyer to confirm.

25

Calendar

26

About this document

The regulatory and legal research for this study drew on the Psiquepol + AI corpus of laws, treaties and monitoring, the platform we use to track regulation and its actors in each country: as of the cutoff date of this document, 84,747 laws and regulations, 484,498 bills and 816,214 parliamentary speeches. For search, source cross-checking and drafting, we used artificial intelligence models.

In the coming days we will publish a companion to this study, “Open and Closed Models: The Dispute That Shapes AI,” on who builds open and closed AI models in the United States, China and elsewhere, which governments and companies sponsor them, where Nvidia stands after Jensen Huang's resignation on November 11, and how the law treats each kind of model.

This is a document of public policy analysis and of the artificial intelligence landscape, intended as input for decision-making. It does not constitute legal, political or financial advice.

27

AI Monitoring

Outside our platform, some official sources restrict access depending on the country or the type of connection. Psiquepol + AI's AI Monitoring brings together the laws, bills and debates cited in this study, with a link to each source and to the profile of each actor, institution and article, and updates them continuously by country or with global coverage. It is available immediately in countries with active monitoring; in other countries, it is activated within 72 hours or your money is refunded.

Subscribe to AI Monitoring

Notes and sources

1. 1.Psiquepol + AI, “AI legal frameworks in 32 countries and the European Union: results,” November 9, 2026, updated November 11, our own measurement over 1,848 cells with a source or a negative search; compute weights from Epoch AI, AI Chip Imports.↩abcdefghijklmnopqrstuv
2. 2.SB 1047 (2023-2024), enrolled text of September 3, 2024, §§ 22602, 22603 and 22606: «Before beginning to initially train a covered model, the developer shall do all of the following» and «Implement the capability to promptly enact a full shutdown» (§ 22603(a)(2)(A)); audit from 2026, § 22603(e); penalties, § 22606(a)(1). Governor Newsom's veto message, September 29, 2024: «the bill applies stringent standards to even the most basic functions - so long as a large system deploys it».↩ab
3. 3.veto message; Federal Register.↩ab
4. 4.SB 53, §§22757.12 and 22757.15.↩ab
5. 5.SB 53, §§22757.13 and 22757.15; California Labor Code, §1107.1; AI Act, art. 55(1)(c).↩ab
6. 6.8 U.S.C. §1184(g)(1).↩
7. 7.Proclamation 11069, 91 FR, September 23, 2026.↩
8. 8.City of Tucson, bulletin of August 8, 2025 and WTOP, August 2025, on the Prince William ruling.↩ab
9. 9.S. 4214, March 25, 2026, and H.R. 9442, June 24, 2026.↩ab
10. 10.University of Melbourne and KPMG, “Trust, attitudes and use of artificial intelligence: A global study 2025,” figure 13.↩abcde
11. 11.SAG-AFTRA, July 9, 2025.↩ab
12. 12.Department of Labor, TEGL 15-25, June 25, 2026.↩ab
13. 13.Human resources and social security plan of the 15th Five-Year Plan, via Science and Technology Daily, July 9, 2026.↩
14. 14.Ray Dalio, Principles for Dealing with the Changing World Order (2021), “Analysis of the World’s Leading Countries,” tables by country, pp. 137 to 166. Book.↩abc
15. 15.Ray Dalio, The Changing World Order: Where We Are and Where We're Going (2020), pp. 301, 311, 312 and 321. Book, from the collection.↩
16. 16.Digital Daily, “금속노조, 총파업 돌입…현대차·한국GM 등 8만5000명 일손 놓았다”, July 15, 2026.↩
17. 17.King & Wood Mallesons, “Data Centres: APAC Regulatory Guide 2026,” Singapore chapter, November 2026.↩ab
18. 18.Government of Singapore, “How Singaporean Workers Are Supported Through the AI Transition”, May 8, 2026.↩
19. 19.Tseng Wen-sheng, chairman of Taipower, before the Economics Committee of the Legislative Yuan, in udn, July 15, 2026.↩ab
20. 20.Taiwan's Artificial Intelligence Basic Act, arts. 15 and 17.↩ab
21. 21.Ballotpedia, November 6, 2026. It covers counties and municipalities, nine are advisory, and outside the large cities its coverage is partial.↩abcdefgh
22. 22.E&E News, August 4, 2026.↩abc
23. 23.Data Center Dynamics, Entergy and Meta and Mississippi Free Press, xAI, March 10, 2026.↩ab
24. 24.Signal Ohio, July 28, 2026.↩ab
25. 25.Atlanta Journal-Constitution, November 2026.↩ab
26. 26.Psiquepol + AI bill registry: Korea, bill 2218836, enacted; Brazil, PL 278/2026, PLP 74/2026, PLP 23/2026, PL 414/2026 and PL 493/2026, and PL 2338/2023, Chamber of Deputies.↩ab
27. 27.Capacity, September 24, 2026.↩abc
28. 28.Monitoring Analytics.↩abc
29. 29.GE Vernova, third-quarter 2026 results, October 28, 2026: «Gas Power equipment backlog and slot reservation agreements grew from 116 to 122 GW; now anticipate reaching at least 130 GW by year-end 2026». The 2031 deliveries and the data center share, from Utility Dive, on the second quarter.↩ab
30. 30.Pew, young adults.↩abc
31. 31.Pew Research Center, “How Americans View AI and Its Impact on People and Society”, September 17, 2025.↩
32. 32.Ipsos, AI Monitor 2026, June 2, 2026, 32 countries, fieldwork March 20 to April 3, 2026, 23,532 adults; question: «Products and services using artificial intelligence make me nervous».↩
33. 33.Merriam-Webster, 2025 word of the year.↩
34. 34.Nataliya Kosmyna et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task”, arXiv, June 10, 2025. Not peer reviewed.↩
35. 35.Statement on AI training.↩
36. 36.Walter G. Stephan and Cookie White Stephan, “An Integrated Threat Theory of Prejudice,” in Stuart Oskamp (ed.), Reducing Prejudice and Discrimination (2000). Book; definitions taken from secondary sources.↩
37. 37.Jakub Złotowski, Kumar Yogeeswaran and Christoph Bartneck, “Can we control it? Autonomous robots threaten human identity, uniqueness, safety, and resources”, International Journal of Human-Computer Studies, 2017.↩ab
38. 38.“The emotional impact of generative AI: negative emotions and perception of threat”, Behaviour & Information Technology, 2024.↩
39. 39.OECD, Employment Outlook 2026, executive summary, July 2026.↩
40. 40.UBS, Global Wealth Report 2023, p. 7: “the wealth share of the global top 1% falling to 44.5%,” with 2022 data.↩
41. 41.Oxfam, global inequality report, January 19, 2026.↩ab
42. 42.Pew Research Center, “Views of children's financial future”, January 9, 2025, spring 2024 survey in 19 countries.↩
43. 43.Ipsos, What Worries the World, July 2026, 25,540 adults in 30 countries, fieldwork June 19 to July 3.↩
44. 44.MoneyVox, September 10, 2025, with the Interior Ministry's tally.↩
45. 45.Euronews, October 3, 2026.↩
46. 46.Carnegie Europe and European Democracy Hub, “The 2025 Surge in Global Protests Is Continuing in 2026”, June 2026.↩
47. 47.ABC News, The Washington Post and Ipsos, October 21, 2026, fieldwork October 9 to 13, 2,545 adults, margin of 2 points; the question about the community went to half the sample, margin of 2.8.↩abc
48. 48.YouGov, May 2026, 2,170 adults.↩
49. 49.Edelman, Trust Barometer Flash Poll: Trust and Artificial Intelligence at a Crossroads, November 18, 2025, p. 7; fieldwork October 17 to 27, 2025, 5,000 people in Brazil, China, Germany, the United Kingdom and the United States. It is the latest measurement: the January 2026 Trust Barometer repeats it.↩
50. 50.ACLED, September 28, 2026.↩
51. 51.Karsten Paul, Hannah Scholl and Klaus Moser, “Thirty years of research on the latent deprivation model”, Frontiers in Psychology 14, 2023, on the latent functions of employment described by Marie Jahoda.↩
52. 52.Liliana Winkelmann and Rainer Winkelmann, “Why Are the Unemployed So Unhappy? Evidence from Panel Data”, Economica, 1998.↩
53. 53.Karsten Paul and Klaus Moser, “Unemployment impairs mental health: Meta-analyses”, Journal of Vocational Behavior 74, 2009.↩
54. 54.Pew Research Center, “How Americans View Their Jobs”, March 30, 2023.↩
55. 55.OpenAI, “GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks”, arXiv, October 2025; April 2026 figures in OpenAI, “Introducing GPT-5.5”, table «GDPval (wins or ties)», measured by OpenAI.↩
56. 56.Milad Mirbabaie, Felix Brünker, Nicholas Möllmann Frick and Stefan Stieglitz, “The rise of artificial intelligence: understanding the AI identity threat at the workplace”, Electronic Markets, 2021.↩
57. 57.Jennifer Logg, Julia Minson and Don Moore, “Algorithm appreciation: People prefer algorithmic to human judgment”, Organizational Behavior and Human Decision Processes 151, 2019.↩
58. 58.Brian Porter and Edouard Machery, “AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably”, Scientific Reports, November 14, 2024: poems are rated worse when the reader believes an AI wrote them.↩
59. 59.Yang, “Perception of AI-induced occupational replacement risk and withdrawal behavior: an identity perspective”, Frontiers in Psychology, 2026.↩ab
60. 60.Anthropic, “Formalizing Fermat's Last Theorem”, September 4, 2026. It verifies Wiles's proof; it is not new mathematics.↩
61. 61.OpenAI, openai/math repository, created October 6, 2026; the README says about 42% of the main results are formalized and that the unformalized ones could contain errors.↩
62. 62.Terence Tao in Proofs and Prompts, “100+ reactions to 100+ solutions”, October 8, 2026.↩
63. 63.Lex Fridman Podcast #503, October 26, 2026.↩abc
64. 64.Pew Research Center, “Globally, More People Expect AI to Cause Job Loss Than Growth”, September 17, 2026, fieldwork February 8 to May 13, 2026, 42,151 adults in 36 countries plus the United States: «The same trend is seen among people with less education and lower household income». The 2025 edition (October 15, 2025) found the same.↩
65. 65.Pew Research Center, “U.S. Workers Are More Worried Than Hopeful About Future AI Use in the Workplace”, February 25, 2025.↩
66. 66.Sarah Lebovitz, Hila Lifshitz-Assaf and Natalia Levina, “To Engage or Not to Engage with AI for Critical Judgments”, Organization Science, 2022.↩
67. 67.Minkyu Shin, Jin Kim, Bas van Opheusden and Thomas Griffiths, “Superhuman artificial intelligence can improve human decision-making by increasing novelty”, PNAS, 2023.↩
68. 68.Abraham Maslow, “A Theory of Human Motivation”, Psychological Review, 1943.↩
69. 69.Andy Jassy, letter to Amazon employees, June 17, 2025.↩ab
70. 70.Tobi Lütke on X, April 7, 2025.↩
71. 71.Mechanize, “Announcing Mechanize, Inc.”, April 17, 2025; TechCrunch, April 19, 2025, which quotes Tamay Besiroglu's post with the goal and the wage figure.↩
72. 72.Daniel Ames and Susan Fiske, “Intentional Harms Are Worse, Even When They're Not”, Psychological Science 24(9), 2013.↩
73. 73.Kurt Gray, Liane Young and Adam Waytz, “Mind Perception Is the Essence of Morality”, Psychological Inquiry 23(2), 2012.↩
74. 74.Gallup, September 15.↩abcde
75. 75.Christos Makridis, “Using AI More Does Not Reassure Workers, Managers Do”, Gallup, September 9, 2026.↩
76. 76.The Harris Poll for INTOO, Employee Mindset Report, fieldwork November 4 to 6, 2025, 1,158 employees, via CPA Practice Advisor, January 8, 2026.↩
77. 77.Gallup, layoffs.↩ab
78. 78.Yochanan Bigman, Desman Wilson, Mads Arnestad, Adam Waytz and Kurt Gray, “Algorithmic discrimination causes less moral outrage than human discrimination”, Journal of Experimental Psychology: General, 2023.↩
79. 79.Jeremy Kahn, interview with Marc Benioff, Fortune, July 30, 2025.↩
80. 80.NBC News, Lester Holt's interview with Satya Nadella, January 30, 2024.↩
81. 81.CBS News, 60 Minutes, Bill Whitaker's report on Jensen Huang, April 28, 2024.↩
82. 82.Milken Institute, Michael Milken's conversation with Jensen Huang, Global Conference, May 2025, transcript, minute 06:40.↩
83. 83.U.S. Senate, Judiciary Subcommittee on Privacy, Technology, and the Law, hearing “Oversight of A.I.: Legislating on Artificial Intelligence,” September 12, 2023, William Dally's answer to Chair Blumenthal.↩
84. 84.European Union AI Act, article 14, paragraphs 1 and 4(b).↩ab
85. 85.Hao-Ping Lee et al., “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers”, CHI 2025.↩
86. 86.Raja Parasuraman and Dietrich Manzey, “Complacency and Bias in Human Use of Automation: An Attentional Integration”, Human Factors 52(3), 2010.↩
87. 87.Sarah Jabbour, David Fouhey et al., “Measuring the Impact of AI in the Diagnosis of Hospitalized Patients: A Randomized Clinical Vignette Survey Study”, JAMA 330(23), 2023, pp. 2275–2284. Randomized clinical vignette study (simulated cases) with 457 hospitalist physicians, nurse practitioners and physician assistants in 13 states, April 2022 to January 2023: baseline accuracy of 73.0%; with predictions from a systematically biased model, 11.3 points lower (95% CI, 7.2 to 15.5), and 9.1 points lower when the model also showed its explanation; with an unbiased model, 2.9 points higher, and 4.4 with explanations.↩
88. 88.Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone, “When combinations of humans and AI are useful: A systematic review and meta-analysis”, Nature Human Behaviour 8, 2024, pp. 2293–2303: 106 experiments with 370 effect sizes, published between January 2020 and June 2023; compared with the better of the two alone, the combination performs worse (Hedges' g = −0.23; 95% CI, −0.39 to −0.07), with losses in decision tasks.↩
89. 89.Nikhil Agarwal, Alex Moehring, Pranav Rajpurkar and Tobias Salz, “Combining Human Expertise with Artificial Intelligence: Experimental Evidence from Radiology”, NBER Working Paper 31422, July 2023, revised August 2026, not peer reviewed: 227 radiologists; “AI assistance does not improve radiologists' average diagnostic quality, even though its predictions are more accurate than 78% of them,” because “they underweight AI predictions.”↩
90. 90.Thomas Dratsch et al., “Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance”, Radiology 307(4), 2023: 27 radiologists, with a simulated AI; the accuracy of the most experienced fell from 82.3% to 45.5%.↩
91. 91.Madeleine Clare Elish, “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction”, Engaging Science, Technology, and Society 5, 2019.↩
92. 92.Edmond Awad et al., “Drivers are blamed more than their automated cars when both make mistakes”, Nature Human Behaviour 4, 2020.↩
93. 93.The Register, September 2, 2025, on what Benioff said on The Logan Bartlett Show, with Salesforce's response.↩
94. 94.Sam Altman, “Three Observations”, February 2025.↩
95. 95.Anthropic, Economic Index, report of September 15, 2025.↩
96. 96.Dwarkesh Patel, interview with Elon Musk, February 5, 2026, minutes 1:15 to 1:16.↩
97. 97.Anthropic, S-1/A of November 10, 2026, amendment no. 2. The valuation is our own calculation: 7,427,991,791 Class A and 5,658,680,800 Class N shares outstanding after the offering and the private placement, at $74.↩abcdef
98. 98.AI Futures Project, “Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All”, November 10, 2026, by Eli Lifland, Brendan Halstead, Daniel Kokotajlo and others.↩ab
99. 99.Naomi Aoki, “The importance of the assurance that ‘humans are still in the decision loop’ for public trust in artificial intelligence: Evidence from an online experiment”, Computers in Human Behavior 114, 2021, with 1,542 long-term care users and their families.↩
100. 100.Barry Staw, Lance Sandelands and Jane Dutton, “Threat Rigidity Effects in Organizational Behavior: A Multilevel Analysis”, Administrative Science Quarterly 26(4), 1981.↩
101. 101.Writer, enterprise AI adoption survey, April 7, 2026, with Workplace Intelligence, fieldwork December 17, 2025 to January 25, 2026, 1,200 executives and 1,200 employees in the United States, the United Kingdom, Ireland, Benelux, France and Germany. Survey by a company that sells AI; the sabotage is self-reported. The March 2025 edition gave 31%, in the United States only.↩
102. 102.Arizona Capitol Times, August 19, 2026.↩ab
103. 103.Eric Hobsbawm, “The Machine Breakers,” Past & Present 1 (1952), at libcom.↩abc
104. 104.E. P. Thompson, The Making of the English Working Class (1963), ch. 14, and Kevin Binfield, Writings of the Luddites (Johns Hopkins University Press, 2004), who dates the movement from March 1811 to April 1817, with his region-by-region summary at Murray State University. The gig mill act (5 & 6 Edward VI c. 22) and the woollen apprenticeship laws were repealed in 1809.↩ab
105. 105.Kirkpatrick Sale, Rebels Against the Future (1995), and Adrian Randall, “Reinterpreting ‘Luddism’”, in Martin Bauer (ed.), Resistance to New Technology (Cambridge University Press, 1995), pp. 57-80.↩
106. 106.Lord Byron in the House of Lords, February 27, 1812, Hansard.↩
107. 107.Murray Edelman, The Symbolic Uses of Politics (1964), p. 6, at archive.org.↩
108. 108.Aina Gallego and Thomas Kurer, “Automation, Digitalization, and Artificial Intelligence in the Workplace: Implications for Political Behavior”, Annual Review of Political Science 25, 2022.↩
109. 109.Thomas Kurer, “The Declining Middle: Occupational Change, Social Status, and the Populist Right”, Comparative Political Studies, 2020.↩ab
110. 110.Nicole Wu, “Misattributed blame? Attitudes toward globalization in the age of automation”, Political Science Research and Methods, 2021.↩
111. 111.Carl Benedikt Frey, Thor Berger and Chinchih Chen, “Political machinery: did robots swing the 2016 US presidential election?”, Oxford Review of Economic Policy, 2018.↩
112. 112.Massimo Anelli, Italo Colantone and Piero Stanig, “Individual vulnerability to industrial robot adoption increases support for the radical right”, PNAS, 2021.↩
113. 113.Gallup, open-ended question to those who oppose a data center in their area, panel of April 1 to 15, 2026, 1,561 opponents. Gallup does not publish the overlap between categories: AI-specific reasons range from 14% if they fully overlap to 27% if they do not overlap.↩
114. 114.Insights & Indicators, “The data center backlash”, which compiles the experiments by Echelon Insights (from 35 against 53 to 27 against 62) and Change Research (from 37 against 51 to 35 against 55). The date of the Echelon experiment differs between sources.↩ab
115. 115.ITIF and Public First, August 26, 2026, fieldwork August 8 to 12, 2,007 adults.↩ab
116. 116.Caltech, Linde Center for Science, Society, and Policy, “AI data centers are not popular, but does it matter?”, with YouGov, December 2025; reading of its figure 3.↩
117. 117.Gallup, “Most Important Problem”, monthly series, October 2026.↩
118. 118.Bureau of Labor Statistics, Current Population Survey, series LNS12000000, employed in October 2026. The 44 million figure is our estimate: it applies Gallup's percentage to total employment.↩
119. 119.Bureau of Labor Statistics, “Union Members, 2025”, February 18, 2026.↩
120. 120.Pew Research Center, April 20, 2021, fieldwork April 5 to 11, 5,109 adults.↩
121. 121.House of Representatives, roll call 75 of March 11, 2021, H.R. 8, Bipartisan Background Checks Act.↩
122. 122.Pew Research Center, November 19, 2018, which compares its 2018 survey (49%) with the 2016 one (39%).↩
123. 123.Vermont, Act 120 of 2014, section 7.↩
124. 124.Public Law 114-216, July 29, 2016, National Bioengineered Food Disclosure Standard.↩
125. 125.NBC News exit poll.↩abcdef
126. 126.NPR, September 15, 2026; Newsweek. Town hall organized by the Future of Life Institute.↩
127. 127.Liz Shuler at the National Press Club, April 28, 2026; National Press Club.↩
128. 128.Semafor, October 6, 2026; Senator Gallego's press release.↩
129. 129.H.R. 10044, AI Tax and Work Protection Act, August 6, 2026.↩
130. 130.Abdul El-Sayed, “AI Under Democracy”, June 2026; November 3 result in Fox, November 4, 2026.↩
131. 131.Dario Amodei, “The Adolescence of Technology”, January 2026.↩abcde
132. 132.Dario Amodei, “Policy on the AI Exponential”, June 11, 2026.↩ab
133. 133.Anthropic, Economic Policy Framework, June 2026.↩ab
134. 134.Sam Altman, “Moore's Law for Everything”, March 16, 2021.↩
135. 135.OpenAI Global Affairs, “Industrial Policy for the Intelligence Age”, April 6, 2026.↩
136. 136.European Union AI Act, article 26(7).↩
137. 137.The Intercept Brasil, December 2, 2024, on PL 2338/2023.↩
138. 138.Psiquepol + AI bills corpus, query of November 11, 2026: 73 bills from 2025 and 2026 in 11 countries whose title combines AI and work. The exceptions are Brazil's PL 2067/2026, 2019/2026 and 5178/2026, Italy's AC 2842 and the bill amending article 439 of Mexico's Federal Labor Law.↩
139. 139.Dario Amodei, “Machines of Loving Grace”, October 2024.↩ab
140. 140.Eva Vivalt et al., “The Employment Effects of a Guaranteed Income,” Quarterly Journal of Economics 141(4), 2026; NBER w32719, revised August 2026.↩
141. 141.Pew Research Center, October 4, 2017.↩
142. 142.Blue Rose Research, December 9, 2025, 12,000 people, via Bharat Ramamurti. Secondary source.↩
143. 143.More in Common, May 13, 2026. It does not publish the sample size.↩
144. 144.Leo XIV, Magnifica Humanitas, May 15, 2026, ¶149 and ¶167.↩
145. 145.Klaus Schwab, “The Fourth Industrial Revolution: what it means and how to respond”, World Economic Forum, January 14, 2016; first published in Foreign Affairs.↩ab
146. 146.Pew Research Center: confidence, Americans and AI 2026, June 17, 2026, fieldwork February 17 to 23, 2026, 5,119 adults («67% of Americans have little to no confidence in the U.S. government to regulate AI effectively. This is slightly up from 62% in 2024»; in the topline, confidence in companies, «not too much» 38% and «none» 21%, 59%, the same as in 2024); jobs, September 16, 2026 and the U.S. table of the global report, fieldwork June 22 to 28, 2026, 3,488 adults (71% expect fewer jobs in 20 years). In August 2024, 64% expected fewer jobs (April 3, 2025).↩
147. 147.Bentley-Gallup, Americans Cool Toward AI.↩abcd
148. 148.Reuters/Ipsos via Allwork.↩
149. 149.Quinnipiac, March 2026.↩ab
150. 150.Gallup via Fortune/AP.↩
151. 151.Gallup, AI indicator.↩ab
152. 152.Gallup, American Job Quality Study.↩abc
153. 153.Pew, use by age.↩
154. 154.Fortune, June 11, 2025 and Tom's Hardware.↩ab
155. 155.Challenger, Gray & Christmas, October 2026.↩ab
156. 156.Challenger, September 2026.↩
157. 157.Klarna, F-1/A and Entrepreneur.↩
158. 158.Duolingo, 10-K.↩
159. 159.Cursor.↩
160. 160.Hosseini and Lichtinger, SSRN.↩
161. 161.Dallas Fed, September 22.↩
162. 162.St. Louis Fed.↩
163. 163.BLS, 2025-35 projections.↩
164. 164.Yale Budget Lab, New York Fed and Dallas Fed.↩
165. 165.BLS, October employment.↩abc
166. 166.Barr and Cook.↩
167. 167.New York Fed, graduates.↩
168. 168.Anthropic, labor market impacts.↩
169. 169.Brookings, capacity to adapt.↩ab
170. 170.Stanford AI Index 2026, public opinion. The figures come from Pew's 2024 survey of experts and the public (April 3, 2025); Pew has not repeated the expert survey.↩
171. 171.Daron Acemoglu and Simon Johnson, Power and Progress, PublicAffairs, 2023.↩
172. 172.World Economic Forum, Future of Jobs Report 2025, press release of January 8, 2025. Survey of more than 1,000 employers in 22 industries and 55 economies.↩abcd
173. 173.Robert C. Allen, “Engels' pause: Technical change, capital accumulation, and inequality in the British industrial revolution”, Explorations in Economic History 46 (4), 2009, pp. 418-435.↩
174. 174.David Autor, David Dorn, Gordon Hanson and Kaveh Majlesi, “Importing Political Polarization? The Electoral Consequences of Rising Trade Exposure”, American Economic Review 110 (10), 2020, pp. 3139-3183.↩
175. 175.Eva Vivalt et al., “The Employment Effects of a Guaranteed Income,” Quarterly Journal of Economics 141(4), 2026, pp. 2675-2734; NBER w32719, revised August 2026: «total individual income fell by around 16 cents». The first version, from July 2024, gave 1.3 hours and 12 cents (Business Wire, July 21, 2025).↩
176. 176.Business Insider, April 2026, on an interview with Nicholas Thompson for The Atlantic.↩
177. 177.The Washington Times, April 17, 2026 and Reason, April 17, 2026.↩
178. 178.Gizmodo, June 2026.↩
179. 179.Alaska Statutes §37.13.145(b).↩
180. 180.China's National Bureau of Statistics via Trading Economics.↩
181. 181.ChinaTalk, “China on AI Job Loss,” April 1, 2026.↩ab
182. 182.State Council of China, guidelines of October 23, 2026, via Reuters on Allwork.↩
183. 183.Law firms Littler and DLA Piper on California SB 951 and Connecticut Public Act 26-15.↩ab
184. 184.New York, A9533.↩
185. 185.S. 3108.↩
186. 186.S. 3108 and H.R. 9352, by Steven Horsford, June 18, 2026.↩
187. 187.Quinnipiac, September 2026.↩
188. 188.New York Civil Service Law, §80(10)(b).↩
189. 189.Mexico's Federal Labor Law, art. 439; Constitution of Brazil, art. 7, XXVII; France's Labor Code, L1233-3 and L1233-4; Spain's Workers' Statute, art. 64.4.d.↩
190. 190.Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update, June 18, 2026, executive summary and section 1. The previous edition (press release of January 15, 2025) gave 4.4% in 2023 and 6.7% to 12% in 2028; the new one revised earlier years downward.↩
191. 191.Lawrence Berkeley National Laboratory, Shehabi et al., 2024 United States Data Center Energy Usage Report, December 2024, pp. 55 to 57. Indirect water does not discount power purchase agreements or each data center's own generation.↩
192. 192.Scientific American, November 4.↩
193. 193.BLS API, CUUR0000SEHF01; Time; Campaigns & Elections; Pew.↩abc
194. 194.Edison Research at SSRS, AI User Metrics, May 2026. Probability sample of adults 18 and older.↩
195. 195.Heatmap.↩
196. 196.Wood Mackenzie, August 2026 projections, via Bloomberg on Energy Connects, August 12, 2026: «committing to about 28% of the 1,066 gigawatts requested for data center projects».↩
197. 197.Earthjustice, October 30, 2026.↩
198. 198.The Register, on docket CL26002535-00, and OPB, March 31, 2026.↩
199. 199.MIT Technology Review.↩ab
200. 200.Denver7/Scripps, governors and data centers, unofficial results.↩abc
201. 201.Texas Tribune.↩ab
202. 202.Data Center Dynamics.↩ab
203. 203.Click2Houston, November 4.↩
204. 204.ABC News 4, Shapiro's executive order.↩
205. 205.ABC7 New York.↩
206. 206.Psiquepol + AI bill registry, 119th Congress bills with “data center” in the title, without duplicates, as of November 8, 2026. Based on Congress records.↩
207. 207.Clerk, vote 312; Politico, via Yahoo.↩
208. 208.Roll Call, September 30, 2026.↩
209. 209.Womble Bond Dickinson, The Bipartisan American Affordability and Jobs Act and C2ES.↩
210. 210.ABC News 4, expansion of the pledge. The figures are the White House's.↩
211. 211.Broadband Breakfast.↩
212. 212.Capacity, January 14, 2026 and Anthropic, February 11, 2026.↩
213. 213.CNBC, Democrats after the election.↩ab
214. 214.Bernie Sanders on X, November 4 and Quartz.↩
215. 215.E&E News, moratorium on Capitol Hill.↩
216. 216.Fox News, March 2026.↩ab
217. 217.Executive Order 14318, July 23, 2025, sections 2(d), 3, 4 and 9.↩abc
218. 218.Department of Energy, selection of four sites for AI data centers, July 24, 2025.↩
219. 219.POWER Magazine, on the selection of July 29, 2026; the final agreement is missing.↩
220. 220.U.S. Army, March 26, 2026. Conditional selection under 10 USC §2667.↩
221. 221.Defense News, October 23, 2025 and Tullahoma News, June 1, 2026.↩
222. 222.Presidential determination under section 303 of the Defense Production Act, April 20, 2026, 91 FR 21931.↩
223. 223.50 U.S.C. §§4511, 4555, 4557, 4558 and 4564, version of the U.S. Code in force as of November 8, 2026, and 15 CFR §700.13. A February 2026 Congressional Research Service report gave September 30 as the expiration date; the date in force is later than that report.↩abc
224. 224.The National, December 5, 2025 and The Japan Times, November 19, 2025. No confirmation that the first phase in Abu Dhabi is already operating.↩ab
225. 225.Executive Order 14365, December 11, 2025, section 8(b).↩
226. 226.H.R. 10119, Protecting Local Control of Data Centers Act, 119th Congress, referred to committee on August 20, 2026.↩
227. 227.France's Urban Planning Code, L300-6-2 and Germany's Energy Efficiency Act, §11.↩
228. 228.HM Treasury, The Data Centre Covenant, October 28, 2026. The definition of a strategic center is still pending.↩
229. 229.NYU Stern, Center for Business and Human Rights, April 15, 2026, which cites Tom's Hardware.↩
230. 230.FCC, Space Bureau, public notice DA 26-113, February 4, 2026, docket SAT-LOA-20260108-00016, and comments by the Secure World Foundation, March 6, 2026. The FCC has not ruled.↩ab
231. 231.Sundar Pichai, third-quarter results, October 28, 2026, and call transcript. The second is a reproduction of the call.↩abc
232. 232.Axios, November 5, 2026, which cites Michael Cembalest (JPMorgan), Krishna Guha (Evercore), the Financial Times Lex column and Derek Thompson, with FactSet data for the stocks.↩abc
233. 233.EIA, Today in Energy, February 20, 2026.↩
234. 234.26 U.S.C. §45Y(d)(4) for the placed-in-service date; the construction-start condition, in Pierce Atwood, on the 2025 law and IRS Notice 2025-42.↩ab
235. 235.Executive Order 14300, May 23, 2025 and White House fact sheet.↩
236. 236.American Nuclear Society, Meta, January 12, 2026, American Nuclear Society, Kairos and Google and World Nuclear News, Crane.↩
237. 237.Data Center Dynamics, Google and Fervo and American Nuclear Society, Helion, June 23, 2026.↩
238. 238.Nicholas Institute, Duke University, February 11, 2025.↩
239. 239.Google, August 4, 2025 and Yahoo News, on the withdrawal in Franklin Township.↩
240. 240.Department of Energy, order 202-26-23 to PJM under section 202(c) of the Federal Power Act.↩
241. 241.Akin Gump, FERC orders of June 18, 2026.↩
242. 242.Utility Dive, on the FERC order of October 9, 2026.↩
243. 243.AP, via Boston Globe.↩ab
244. 244.National Nurses United, July 22, 2026.↩
245. 245.JLARC.↩ab
246. 246.Semafor, October 15, 2026. Survey by Fabrizio Ward and Impact Research.↩abc
247. 247.Virginia Mercury, August 5, 2026.↩ab
248. 248.AEP Ohio (part), Georgia PSC, Texas SB 6, Virginia via SELC; DCK, IBJ.↩
249. 249.WUSF and News Service of Florida, November 4, 2026.↩
250. 250.Florida §366.043; Idaho §61-335; South Dakota §49-50-2; Delaware, tit. 26, §317; Texas Utilities Code §39.170.↩
251. 251.CNBC, analysis of November 6.↩abcd
252. 252.China's National Energy Administration, China Electricity Market Development Report 2025, June 18, 2026: «新增发电装机容量5.42亿千瓦» (542 million kilowatts of new capacity). The eleven-month figure, about 445 gigawatts, came from the South China Morning Post.↩ab
253. 253.The Star, with data from the National Energy Administration, January 28, 2026; CREA and Global Energy Monitor; and EIA, May 2026, with data from the IAEA.↩
254. 254.International Energy Agency, Energy and AI, energy demand, base case.↩
255. 255.State Council of China, “Opinion on the in-depth implementation of the ‘AI+’ action,” Guofa (2025) 11, August 21, 2025, section I and section II, item (4) 1.↩abcde
256. 256.MIT Technology Review, “China built hundreds of AI data centers to catch the AI boom. Now many stand unused,” March 26, 2025. The 80% figure is from the Chinese outlets Jiazi Guangnian and 36Kr, cited by the magazine; the 500-project figure is from KZ Consulting.↩ab
257. 257.AI Kill Switch and Data Centre Control Bill 2026, Parliament of Australia; draft resolution by François Ruffin, National Assembly, and BT-Drs. 21/8111, Bundestag.↩
258. 258.The Next Web, on Axios.↩abc
259. 259.Delaware, tit. 9, §401; New Jersey §34:11-56.102; Virginia §15.2-2209.5.↩
260. 260.Pew Research Center, “How opinions and use of AI differ by age”, June 17, 2026, survey of 5,119 adults conducted February 17 to 23, 2026. The 49% answers question CHATUSEMOD of the questionnaire, whether respondents ever use an AI chatbot like ChatGPT, Gemini or Copilot; each brand was asked about separately, to half the sample.↩
261. 261.Pew, October 23, 2026.↩
262. 262.Louisiana, Microsoft, Pennsylvania; Axios.↩
263. 263.larson.house.gov; Breaking Defense.↩ab
264. 264.goldmansachs.com.↩
265. 265.Meta, August 12, 2026; Axios, Axios.↩
266. 266.Nvidia, Apple; Microsoft, Mississippi.↩
267. 267.FEC, Leading the Future, as of October 14.↩
268. 268.San Francisco Standard, October 31, 2026.↩abc
269. 269.FEC, committee C00916114 and Fortune, November 4, 2026, for the network.↩ab
270. 270.Fortune, November 4, 2026 and SF Standard, October 31, 2026.↩
271. 271.Fortune, November 4, 2026, with FEC data.↩ab
272. 272.California Secretary of State, district 11, unofficial count and Semafor, November 5, 2026.↩
273. 273.Anthropic, donation to Public First Action.↩
274. 274.The New York Times, via AI Weekly, October 16, 2026 and GV Wire.↩
275. 275.AP-NORC, October 8, 2026, 2,140 adults, September 24 to 28, margin of 2.9 points.↩
276. 276.Dwight D. Eisenhower's farewell address, January 17, 1961, National Archives.↩
277. 277.Rebecca U. Thorpe, The American Warfare State: The Domestic Politics of Military Spending, University of Chicago Press, 2014, per the review by Benjamin O. Fordham, Michigan War Studies Review, October 30, 2014 and the University of Washington interview, June 30, 2014.↩ab
278. 278.Office of Local Defense Community Cooperation, Defense Spending by State, Fiscal Year 2024, November 2025, and its spreadsheet by county and contractor.↩ab
279. 279.Fiscal 2025 10-K reports on SEC EDGAR, properties section: Lockheed Martin, General Dynamics, Northrop Grumman and Huntington Ingalls. We calculate the districts with the Census Bureau's geography API on the approximate location of each plant.↩
280. 280.Lockheed Martin, fact sheet, May 2026; Lockheed Martin, economic impact of the F-35, accessed November 7, 2026; and, for the B-21, the Northrop Grumman press release of November 2023 as reported by the press. The figures are published by the companies and have no independent verification by state.↩
281. 281.James Fallows, “The Tragedy of the American Military,” The Atlantic, January-February 2015, via Ken Opalo, January 3, 2015; for the B-1, Gordon Adams cited by Rolling Stone.↩
282. 282.Kenneth R. Mayer, The Political Economy of Defense Contracting, Yale University Press, 1991, per the Foreign Affairs review.↩
283. 283.Larry M. Bartels, “Constituency Opinion and Congressional Policy Making: The Reagan Defense Buildup,” American Political Science Review 85(2), 1991, doi 10.2307/1963169, abstract and p. 464.↩
284. 284.Gallup, “Military and National Defense” series, measurement of February 1 to 20, 2024, the most recent.↩
285. 285.Congressional Research Service, IN12641 (January 20, 2026) and IN12653 (February 10, 2026), on Public Law 119-60. The 65-year streak is given by law firms such as Holland & Knight.↩
286. 286.Department of War, Comptroller, Fiscal Year 2027 Budget Request Overview Book, figure 1.1 and AI section. The press release of April 21, 2026 rounds the request to $1.5 trillion.↩ab
287. 287.Military.com.↩
288. 288.Congressional Record and press release from Senators Collins and Snowe, for the 44 states; Larry Bartels, American Political Science Review, 1991, doi 10.2307/1963169.↩
289. 289.2024 letter, 2025 letter, Murkowski; Inside Climate News, Sidley.↩
290. 290.Clerk, vote 86; P.L. 117-169, public fact not reopened.↩
291. 291.Department of Defense Chief Digital and Artificial Intelligence Office (CDAO), press release of July 14, 2025, and Built In.↩ab
292. 292.OpenAI, “Introducing OpenAI for Government,” June 16, 2025.↩abc
293. 293.JobsOhio, on Anduril's Arsenal-1 in Pickaway County, with 4,008 jobs committed through 2035 and $310 million in state support.↩
294. 294.Michael Brenes and William D. Hartung, “Private Finance and the Quest to Remake Modern Warfare,” Quincy Institute, June 3, 2024.↩
295. 295.Roberto J. González, Costs of War, Brown University, April 17, 2024.↩
296. 296.Third-quarter 2026 LD-2 reports, via AI Weekly and GadgetReview. The amounts match the LD-2 forms of Nvidia, Anthropic and OpenAI.↩
297. 297.Third-quarter 2026 LD-2 forms: Meta, Amazon and Google. Extracted on November 8, 2026; the Senate Office of Public Records is not responsible for the data or for derived analyses once extracted.↩
298. 298.Yaqiu Wang and Nicholas Bequelin, Project Syndicate, September 4, 2026.↩
299. 299.Xinhua, January 22, 2017, creation of the Central Commission for Integrated Military and Civilian Development, chaired by Xi Jinping; State Council, New Generation Artificial Intelligence Development Plan, 国发〔2017〕35号, July 20, 2017, section “加强人工智能领域军民融合”; Department of Defense, annual report to Congress on China's military power, December 23, 2025, pp. 62 and 66.↩
300. 300.Shannon Vaughn, Foreign Policy Research Institute, September 10, 2025.↩
301. 301.Goldman Sachs, “Generative AI could raise global GDP by 7%,” April 5, 2023, on the report by Joseph Briggs and Devesh Kodnani.↩
302. 302.McKinsey Global Institute, “The economic potential of generative AI: The next productivity frontier,” June 14, 2023.↩
303. 303.Daron Acemoglu, “The Simple Macroeconomics of AI”, NBER 32487, published in Economic Policy, 2025.↩
304. 304.METR, February 24, 2026: task time with AI changed by -18% (interval from -38% to +9%) for returning programmers and by -4% (from -15% to +9%) for new ones; METR calls the signal unreliable because of participant and task selection and considers the estimate a lower bound on the effect. Earlier experiment: METR, July 10, 2025, 16 experienced programmers and 246 tasks from February to June 2025, with tools that predate agents: with AI, tasks took 19% longer (from +2% to +39%).↩
305. 305.Emerson Murphy-Hill, Jenna Butler and Alexandra Savelieva, “Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI”, arXiv, July 1, 2026. The effect held throughout the study's four months. The authors use merged pull requests as a proxy for output and caution that they are not the same as the value they deliver.↩
306. 306.Ivan Yotzov, José María Barrero, Nicholas Bloom, Steven J. Davis et al., “Firm Data on AI,” NBER w34836, February 2026, revised October 2026, not peer reviewed. It replaces MIT NANDA's 2025 preliminary report (95% with no measurable effect, via Fortune).↩
307. 307.Fortune, October 30.↩abcd
308. 308.Reuters, November 4.↩abc
309. 309.BEA, national accounts, table 1.1.1 (series A191RL) and tables 4.2.5 and 4.2.1 (series B852RC and B852RL). The contribution to growth is our own calculation.↩abcde
310. 310.Fox Business, Trump.↩
311. 311.Jan Hatzius at the Atlantic Council, February 2026, via Gizmodo and Tom's Hardware.↩
312. 312.St. Louis Fed.↩
313. 313.ING, James Knightley, August 20, 2026, which also cites Jason Furman's figure.↩
314. 314.Goldman Sachs (Rindels and Mericle) via Axios, August 12, 2026.↩
315. 315.BLS CES, seasonally adjusted series CES0000000001, CES6562000001, CES5000000001, CES5051800001 and CES6054150001, read from the BLS public API after the November 6, 2026 release (October 2025 to October 2026).↩
316. 316.Fox Business, Moody's and Minneapolis Fed.↩
317. 317.Dean Baker, CEPR, “Five takeaways from the third quarter GDP report”, Substack version.↩
318. 318.Third-quarter 2026 earnings releases and 8-K filings of Microsoft, Alphabet, Meta and Amazon at the SEC: operating cash flow minus capital expenditure (our own calculation). Microsoft excludes finance leases; including them, Fortune estimates its spending at $45.2 billion and free cash flow would fall to about +12,800.↩abc
319. 319.Third-quarter 2026 10-Q forms from Amazon, Alphabet, Meta and Microsoft. Amazon's debt is at face value.↩abcd
320. 320.Meta, press release of October 26, 2026, 8-K and 424B5 from Meta; Microsoft 8-K, November 4; Oracle 8-K, November 4; Alphabet FWP, November 2.↩
321. 321.CNBC, September and October statement.↩
322. 322.Yahoo Finance.↩
323. 323.Reuters, November 5.↩abc
324. 324.24/7 Wall St., November 6.↩ab
325. 325.Fox Business, experts.↩
326. 326.Yahoo Finance, November 6, Fox Business and StockTitan; the weekly tally is our own calculation against the October 30 close.↩ab
327. 327.Reuters via BNN Bloomberg, November 6, 2026, with data from Truist, Vanda and ICE.↩
328. 328.Fortune, July 26, 2026. List prices. The DeepSeek V4 Pro price that Fortune published, about $0.87, is no longer current; the current one comes from DeepSeek's pricing page [518].↩abcdefg
329. 329.TechCrunch, “ChatGPT reaches 900M weekly active users”, February 27, 2026.↩
330. 330.Sundar Pichai, message on Alphabet's third-quarter 2026 results, October 28, 2026. The infographic on the same page gives it as monthly active users, a different measure from ChatGPT's weekly figure.↩
331. 331.Aaron Chatterji et al., “How People Use ChatGPT”, NBER Working Paper 34255, September 2025.↩
332. 332.Anthropic, Anthropic Economic Index, March 2026 report, March 24, 2026, on Claude.ai conversations.↩
333. 333.Microsoft, “The continued state of global AI diffusion in 2026”, September 21, 2026.↩
334. 334.U.S. Chamber of Commerce, 2026 Empowering Small Business, October 2, 2026, with Teneo, 3,732 small businesses, fieldwork July 23 to August 6, 2026: «2/3 of U.S. small businesses (66%) now use AI, up 8 points from 2025 (58%)».↩
335. 335.Alexander Bick, Adam Blandin and David Deming, Generative AI Adoption Tracker, in FRED, “Does generative AI save time at work?”, August 27, 2026 and the series RPSGENAITSALL: «grew from 28.2% in Q3 2024 to 39.2% by Q2 2026» and, over «roughly the same stretch», «grew from 1.6% to 2.2%». Self-reported answers from adults aged 18 to 64. Their earlier estimate, 5.4% of users' hours, is from February 2025.↩
336. 336.Mikhail Bogdanov, JetBrains, “AI Coding Agents: Adoption Trends”, August 18, 2026, with data from the Developer Ecosystem Survey 2026.↩
337. 337.Mikhail Bogdanov, JetBrains, “How Much Code Do Developers Really Let Agents Write?”, August 26, 2026.↩
338. 338.SemiAnalysis, “Claude Code is the Inflection Point”, February 5, 2026.↩
339. 339.Anthropic, “Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation”, February 12, 2026.↩
340. 340.OpenAI, “Codex for every role, tool, and workflow”, June 2, 2026.↩
341. 341.Lovable, “A first look at the build economy”, June 9, 2026.↩
342. 342.GitHub, Octoverse 2026, October 28, 2026, data from October 1, 2025 to September 30, 2026.↩
343. 343.Google DeepMind, “Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad”, July 21, 2025.↩
344. 344.Google DeepMind, “Gemini achieves gold-medal level at the International Collegiate Programming Contest World Finals”, September 17, 2025.↩
345. 345.Google DeepMind, “AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms”, May 14, 2025.↩
346. 346.Royal Swedish Academy of Sciences, press release for the 2024 Nobel Prize in Chemistry, October 9, 2024.↩
347. 347.RCSB Protein Data Bank, our count of experimental entries in its search API, November 10, 2026: 262,786 in total and 17,599 released in 2025.↩
348. 348.OpenAI, “math” repository, README and change history, accessed November 10, 2026.↩ab
349. 349.Thomas Bloom, post of October 8, 2026 on erdosproblems.com.↩ab
350. 350.arXiv 2601.07421, January 2026, on Erdős problem 728.↩
351. 351.Thomas Bloom, post on the unit distance conjecture on erdosproblems.com, May 2026.↩
352. 352.Quanta Magazine, “AI Has Solved One of Math's $1 Million Millennium Prize Problems”, September 8, 2026; CNBC, “OpenAI claims to have solved the 90-year-old Navier-Stokes math problem in 88 hours”, September 9, 2026; The Tufts Daily, “Mathematicians still checking the Navier-Stokes proof that OpenAI claims to have solved”, September 2026.↩ab
353. 353.Al Jazeera, “AI makes discoveries in mathematics, computer science, physics, OpenAI says”, October 27, 2026.↩ab
354. 354.Google DeepMind, arXiv 2601.22401, January 2026, review of 700 Erdős problems.↩
355. 355.Veronica Hernström et al., “Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI)”, The Lancet Digital Health, March 2025.↩
356. 356.Kristina Lång et al., “Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI)”, The Lancet Oncology, August 2023.↩
357. 357.Jessie Gommers et al., “Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study”, The Lancet, January 2026, vol. 407, pp. 505-514.↩
358. 358.Ilan Price et al., “Probabilistic weather forecasting with machine learning”, Nature, December 2024, and Google DeepMind, December 4, 2024.↩
359. 359.Grey Nearing et al., “Global prediction of extreme floods in ungauged watersheds”, Nature, March 2024, and Google Research, March 20, 2024.↩
360. 360.White House, Winning the Race: America's AI Action Plan, July 2025.↩
361. 361.State Council of China, opinion on the "AI+" action, Guofa [2025] No. 11, published August 26, 2025.↩
362. 362.Leopold Aschenbrenner, Situational Awareness: The Decade Ahead, June 2024.↩
363. 363.Sam Altman, OpenAI livestream of October 28, 2025, via eWeek.↩
364. 364.Sam Altman at DevDay 2026, via The Next Web, September 29, 2026.↩
365. 365.Dan Hendrycks, Eric Schmidt and Alexandr Wang, Superintelligence Strategy, March 2025.↩
366. 366.Anthropic, S-1/A of November 2, 2026.↩abcde
367. 367.METR, “Anthropic AI R&D Acceleration Assessment,” November 5, 2026, with the evaluation from August 12 to October 29, and review of the R&D section of the August risk report, which cites the definition of the threshold in the Responsible Scaling Policy. Anthropic had the right to redact non-public information. The threshold text in the prospectus is new in the November 2 amendment and was not in the October 26 S-1.↩abcd
368. 368.Anthropic, “Measurements for understanding the pace of AI development inside frontier labs”. The classification of who directs each task is done by a Claude-based judge, and the agent count covers only the most widely used internal platform.↩
369. 369.Sam McCandlish on X, November 5, 2026 and Daniel Kokotajlo on X, November 3 and 5, 2026.↩ab
370. 370.Anthropic, “Claude Fable 5.2 and Claude Mythos 5.2”, November 9, 2026.↩
371. 371.OpenAI, “Covert underperformance on safety research tasks”, November 4, 2026.↩
372. 372.Anthropic, “An update on independent evaluators”, October 15, 2026.↩
373. 373.S. 5493, Ban Artificial Superintelligence Act of 2026, GovInfo, introduced September 23, 2026 and referred to the Committee on Commerce, Science, and Transportation.↩ab
374. 374.AP via PBS NewsHour, September 23, 2026.↩ab
375. 375.Future of Life Institute, “Statement on Superintelligence”, October 2025; signatures counted on November 7, 2026.↩
376. 376.TechRepublic, September 2026.↩
377. 377.Anthropic, “Claude Gov models for U.S. national security customers,” June 2025.↩
378. 378.Anthropic, “Expanding our partnership with GSA”, October 27, 2026.↩
379. 379.Arvind Narayanan and Sayash Kapoor, “AI as Normal Technology”, Knight First Amendment Institute, April 15, 2025.↩
380. 380.Will Rinehart, “China's AI strategy: adoption over AGI”, American Enterprise Institute, May 8, 2025.↩
381. 381.Matthew Johnson, “AGI Has Quietly Become Central to Beijing's AI Strategy”, China Brief, Jamestown Foundation, October 2025.↩
382. 382.15th Five-Year Plan (2026-2030), presented to the National People's Congress on March 5, 2026, via AI Safety in China, March 27, 2026.↩
383. 383.Jeffrey Ding, Technology and the Rise of Great Powers, Princeton University Press, 2024.↩ab
384. 384.Stanford AI Index 2026, economy.↩
385. 385.Stanford AI Index 2026 via The Next Web. Stanford's research chapter page does not give the percentages.↩abcde
386. 386.Xinhua, December 26, 2025 and Xinhua, December 29, with data from the National Development and Reform Commission and the Ministry of Finance.↩
387. 387.DealStreetAsia, Beijing fund and Yicai, Guangzhou fund.↩
388. 388.Our own conversion at the approximate exchange rate of the date: 7.05 yuan per dollar in December 2025 and 6.7 in 2026 (CurrencyBeacon, September 9, 2026: 6.708).↩ab
389. 389.DeepSeek, “Introducing DeepSeek-V4.1-Flash,” September 10, 2026, API pricing, accessed November 9, 2026, and weights on Hugging Face, MIT license.↩abc
390. 390.Reflection, “Introducing Beam,” October 5, 2026, with the results table reported by the company; “Building Frontier Open Intelligence,” August 9, 2025, for the financing; weights on Hugging Face, Apache 2.0 license.↩
391. 391.eWeek and Forkast, with other cuts of the same platform.↩
392. 392.Stanford AI Index 2026, research and development.↩
393. 393.CSIS, “NVIDIA's Jensen Huang on Securing American Leadership on AI”, transcript of the December 3, 2025 conversation.↩
394. 394.Research Information, “China extends lead in Nature Index global research rankings”, June 10, 2026.↩
395. 395.Anthropic, February 23, 2026.↩
396. 396.NSA, CISA and FBI, advisory AA26-251A, September 8, 2026.↩abc
397. 397.China's Ministry of Commerce, September 9, 2026, via AP on CTV News and National Law Review.↩abcd
398. 398.Anthropic, October 12, 2026.↩
399. 399.Council on Foreign Relations, April 29, 2026 (Chris McGuire, Michael Horowitz and Jessica Brandt).↩abcd
400. 400.Cyfirma, Threat Horizon After Black October. “Black October” is an informal label for independent incidents.↩
401. 401.Wikipedia, “PHP2Shell”, with W3Techs for PHP's market share.↩ab
402. 402.Microsoft Threat Intelligence, October 21, 2026. Also The Record, October 22, 2026.↩abc
403. 403.Axios, October 28, 2026 (Sam Sabin).↩ab
404. 404.Google, “Gemini 4 Carbon: our next frontier model,” October 28, 2026.↩
405. 405.CyberScoop, Notebookcheck and InfoWorld.↩ab
406. 406.Binance, preliminary report, November 3, 2026.↩abc
407. 407.FBI, alert of October 31, 2026. North Korea's response, in SecurityWeek, November 4, 2026.↩abc
408. 408.Epoch AI, How do Chinese AI companies make money?.↩
409. 409.CNN, October 30, 2026.↩ab
410. 410.Bloomberg data via Free Press Journal.↩
411. 411.Morgan Stanley via investingLive; CXMT in The Standard.↩
412. 412.Excitement and nervousness: Ipsos, AI Monitor 2026, pp. 14 to 16, 32 countries, fieldwork March 20 to April 3, 2026: China, 83% excited and 40% nervous; United States, 33% and 64%. The 2025 edition, of 30 countries, via Stanford, was cited before. Jobs and trust: Ipsos, 2025, and Pew, spring 2025 Global Attitudes Survey (25 countries), in Stanford AI Index 2026, chapter 9, figures 9.1.6 and 9.3.2 and findings 4 and 8. In its 2026 global survey, Pew asked the trust-in-own-country question in only 12 countries, without the United States (report of September 17, 2026, p. 32); its national survey of February 2026 gives 67% with little or no confidence in the government (note 823).↩abcde
413. 413.Indmoney.↩
414. 414.CBS, Senate, CBS, House, NBC and CBS exit polls.↩ab
415. 415.Pew, election issues.↩
416. 416.Pew, Trump approval.↩
417. 417.Newsweek.↩
418. 418.CNBC, Texas.↩
419. 419.Boston Globe.↩
420. 420.White House, Executive Order 14277, “Advancing Artificial Intelligence Education for American Youth,” April 23, 2025, sections 2 and 4 to 8; White House, June 30, 2025, more than 60 signatories of the pledge. The figure of more than 135 signatories in September 2025 is from Independent Women, September 12, 2025.↩
421. 421.Xinhua, December 2, 2024, on the notice from the General Office of the Ministry of Education to strengthen AI education in primary and secondary schools. For Beijing: Beijing News, March 2025, 2025-2027 plan with at least 8 hours per school year, and IT之家, September 2025, with CCTV data on more than 1,400 schools.↩
422. 422.SEC; Mitchell et al., Ann Intern Med, doi 10.7326/M20-5665; San Loh et al., JAMA Netw Open, doi 10.1001/jamanetworkopen.2026.3885.↩ab
423. 423.CBS News, July 2025 and Entrepreneur: Microsoft, $12.5 million over five years; OpenAI, $10 million, $2 million of it in technical support and compute credits; Anthropic, $500,000 in the first year.↩
424. 424.aft.org; AMA, AICPA, ABA 512.↩
425. 425.Greg Toppo, The 74, August 4, 2025, republished by Route Fifty.↩
426. 426.Fleischman et al., BMJ 354:i4189; OIG, special fraud alert; Novartis 6-K, DOJ.↩
427. 427.Margalit and Raviv, British Journal of Political Science, DOI 10.1017/s0007123425101282.↩
428. 428.Engadget; Outlook Business.↩
429. 429.Nextgov, Nextgov, protest, Nextgov, 2026.↩
430. 430.University of Melbourne, survey of 48,340 people in 47 countries, in Stanford AI Index 2026, chapter 9, figure 9.1.10.↩
431. 431.David C. Krakauer, “Competitive and Complementary Tools”, arXiv preprint, July 2026; earlier, “Will A.I. Harm Us? Better to Ask How We'll Reckon with Our Hybrid Nature”, Nautilus, 2016.↩
432. 432.StarTalk, “Why We Haven't Reached Superintelligence, with David Krakauer”, October 10, 2026.↩
433. 433.Krzysztof Budzyń et al., “Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study”, The Lancet Gastroenterology & Hepatology, vol. 10, October 2025, pp. 896-903.↩
434. 434.OpenAI, decision on Cursor, August 28, 2026.↩ab
435. 435.Brookings, Mark MacCarthy, March 12, 2026.↩ab
436. 436.Claude, deprecations.↩
437. 437.Regulation (EU) 2019/1150, art. 4(1); Regulation (EU) 2023/2854, art. 25(2)(a); China's Anti-Monopoly Law, art. 22.↩
438. 438.eWeek and Retail Insight Network.↩
439. 439.startups.aws.com; aboutamazon.com; Amazon, 8-K; Alphabet, 8-K.↩ab
440. 440.Shopify; Salesforce; Apple.↩
441. 441.No credits program publishes the retention or the subsequent revenue of the startups it supports.↩
442. 442.Anthropic S-1.↩
443. 443.Menlo; Claude; openai.fund and the Form D of OpenAI Startup Fund II, for $400 million, from August 2026; Meta; Google; the ratios are our own calculation with the amounts from AWS and from each fund.↩
444. 444.Reuters, The Decoder and The Next Web, on documents that are not public.↩
445. 445.SpaceX prospectus.↩
446. 446.15 U.S.C. §13 and FTC, 6(b) study.↩
447. 447.whitehouse.gov.↩
448. 448.Decision of the State Administration for Market Regulation (SAMR) of April 10, 2021, in the press release reproduced in full by The Paper. The document on samr.gov.cn no longer responds.↩
449. 449.SAMR, opening of the investigation into Nvidia, December 9, 2024, and preliminary conclusion of September 15, 2025. They reproduce SAMR's announcements. The entity list: Federal Register, 90 FR 44496, in force since September 12, 2025.↩
450. 450.FTC, staff report on AI partnerships and investments, January 2025.↩
451. 451.Amazon, Form 10-Q as of September 30, 2026. The sum and the 43% are our own calculations on $92.5 billion in preferred shares, $146.2 billion in notes and pre-tax income of $147.56 billion.↩
452. 452.Sequoia, “America's Open-Model Paradox”, with data from the ATOM Project; and The Decoder, Stanford analysis. The ATOM Project promotes American open models.↩ab
453. 453.Verasight, December 2025, per press reports; University of Michigan, consumer survey, preliminary November results, November 6, 2026, open-ended responses with no percentage. Gallup, Pew, AP-NORC, Quinnipiac and YouGov did not ask about a bubble in 2026.↩
454. 454.BoE; Nov. 2025, May 2026.↩
455. 455.CNBC, November 6, 2026, with data from Deutsche Bank, and Bloomberg, October 15, 2026, with data from ICE.↩ab
456. 456.Quartz and The Next Web.↩
457. 457.CNBC, October 29, 2026, on Microsoft's earnings call.↩
458. 458.Reuters via BNN Bloomberg, November 6, 2026.↩ab
459. 459.Bloomberg, November 6, 2026.↩ab
460. 460.Warren letter, warren.senate.gov, FedScoop, on Sarah Friar's remark.↩
461. 461.H.Res. 1642, 119th Congress, introduced October 29, 2026; cosponsors and endorsements in the authors' press release.↩
462. 462.DeSantis on X, November 2025 and June 2026, per Storyboard18 and the Washington Examiner; I did not see the original posts. Punchbowl News on Republican lawmakers.↩
463. 463.whitehouse.gov; Federal Register.↩
464. 464.TechCrunch, November 8, 2025, on the letter of October 27, 2025, and Senator Warren, November 18, 2025.↩
465. 465.12 U.S.C. §343(3)(B); Executive Order 14318, section 3; 50 U.S.C. §4531.↩
466. 466.Anthropic, SEC.↩
467. 467.The September 2025 letter of intent (Nvidia; DCD and IFRE); its replacement by $30 billion in OpenAI's March 2026 round, per Bloomberg (Nvidia did not issue a press release); and the guarantees of up to $105 billion in note 10 of Nvidia's 10-Q as of July 26, 2026.↩
468. 468.The Register, November 4, 2025 and Noah Smith, Noahpinion.↩ab
469. 469.10-K forms at the SEC: Alphabet, 2025, Meta, 2025, Amazon, 2025, Microsoft, fiscal year 2026 and Oracle, fiscal year 2026, property and equipment notes.↩ab
470. 470.CNBC, November 11, 2025, which cites Burry's post on X.↩
471. 471.theCUBE Research, “Resetting GPU depreciation”. Industry analysis.↩
472. 472.Federal Reserve, “Banking Panics of the Gilded Age”, Federal Reserve History.↩
473. 473.National Development and Reform Commission (NDRC), press conference of November 27, 2025.↩
474. 474.21st Century Business Herald, September 29, 2026, with figures from the China Academy of Information and Communications Technology (CAICT), Inspur and the National Data Administration.↩
475. 475.International Energy Agency, Key Questions on Energy and AI, April 16, 2026.↩ab
476. 476.Tom's Hardware, on the gas turbine shortage.↩
477. 477.Lawrence Berkeley National Laboratory, Queued Up: 2026 Edition, June 2026.↩
478. 478.PJM, results of the 2028/2029 capacity auction, July 14, 2026. The previous auction, for 2027/2028, settled at the cap of $333.44 (PJM, December 17, 2025).↩
479. 479.MOFCOM and the General Administration of Customs, announcement No. 58 of 2026, November 9, which keeps announcements 55 to 58 and 61 and 62 of 2025 suspended from November 11, 2026 to January 10, 2027. The tariff, via BusinessWorld and Reuters.↩abc
480. 480.China's Ministry of Commerce, announcement No. 70 of 2025, via Mining Technology and Fastmarkets.↩
481. 481.Wall Street Journal, August 18, 2026, via Crypto Briefing; the ratio of 2.8 times is our own calculation on those figures.↩
482. 482.The Information, June 16, 2026, via The Next Web.↩
483. 483.Financial Times, October 8, 2026, via Bloomberg Government.↩
484. 484.Sam Altman on X, November 6, 2025, via TechCrunch.↩
485. 485.Financial Times, via The Irish Times, February 27, 2026, and PYMNTS.↩
486. 486.The Next Web, on the round closed on March 31, 2026.↩
487. 487.Sam Altman, dinner with reporters in August 2025, per The Verge, via PC Gamer.↩
488. 488.S&P Dow Jones Indices, “In the Shadows of Giants”.↩
489. 489.Nvidia's weight in the S&P 500 as of October 10, 2026, per Slickcharts.↩
490. 490.Gallup, April 2025 survey and April 2026 survey.↩
491. 491.International Monetary Fund, Global Financial Stability Report, October 13, 2026.↩
492. 492.Bank for International Settlements, Annual Economic Report, June 28, 2026, via New Straits Times.↩
493. 493.Bloomberg, November 6, 2026, with Cleanview data, via Indianapolis Business Journal.↩
494. 494.WTOC, November 5, 2026.↩
495. 495.Wall Street Journal, November 5, 2026, via Techmeme.↩
496. 496.Bloomberg, January 22, 2026 and Axios, June 15, 2026.↩
497. 497.European Commission, memorandum of understanding on AI gigafactories.↩
498. 498.Hugging Face, security notice, July 16, 2026, and technical timeline, July 27.↩
499. 499.OpenAI, technical report on the incident, August 26, 2026, pp. 4, 17 and 19 to 21, and joint statement with Hugging Face, July 21.↩abc
500. 500.METR and Redwood Research, independent investigation of the incident, August 26, 2026. Its scope was set by OpenAI and did not cover the attacks on OpenAI's own infrastructure, according to the New York Times.↩abc
501. 501.CNBC, August 26, 2026, for the incident; Wikipedia, for the reactions of Dan Guido and Jake Williams and the “Pacing the Frontier” letter. The subpoena is confirmed by the press release from Attorney General Rob Bonta, October 1, 2026. Definition of frontier model and critical incident: SB 53, §22757.11.↩abc
502. 502.Anthropic, “Agentic Misalignment: How LLMs could be insider threats”, June 20, 2025. These are simulations built to leave the model with no way out, and their critics point out that this exaggerates the likelihood of the behavior.↩ab
503. 503.Anthropic, “Investigating incidents in our cybersecurity evaluations”, July 30, 2026; Dark Reading, on the Meta incident and the disagreement with Irregular.↩
504. 504.AISI, October 23, 2026 and, for the figures in the August 4 report, Al Jazeera, August 5, 2026.↩ab
505. 505.SecurityWeek, November 6, 2026, on OpenAI's report.↩
506. 506.“Beyond the Binary: A nuanced path for open-weight advanced AI”, 2026. Not peer reviewed.↩ab
507. 507.Interesting Engineering and Crypto Briefing, on Google DeepMind's announcement of November 6, 2026.↩
508. 508.OpenAI, “A thousand fewer paths to root,” November 4, 2026.↩
509. 509.BleepingComputer, November 2, 2026, on the CrowdStrike report.↩
510. 510.BleepingComputer, October 31, 2026, on Chainalysis's notice.↩
511. 511.ANSSI, November 9, 2026, on the attribution to Shahid Kaveh, of the Revolutionary Guard. The summoning of the ambassador and the arrest warrants, in The Record, November 9, 2026.↩abc
512. 512.CBS News and AP, November 4, 2026, on the EFS press release and Ahold Delhaize's results.↩ab
513. 513.AFP, via The Star, November 2, 2026, and PBS News, November 2, 2026.↩
514. 514.Business Today, November 3, 2026 and inkl, on SBI's notice to the exchanges.↩
515. 515.YourStory, November 2026 and Deccan Chronicle, on the CERT-In advisory.↩
516. 516.AP, via SecurityWeek, November 6, 2026.↩
517. 517.H.R. 9917, AI Kill Switch Act, July 23, 2026; Roll Call.↩
518. 518.Nextgov, September 10, 2026.↩
519. 519.Tech Policy Press.↩
520. 520.Scott Bessent on CNBC, September 21, 2026, and before the House Financial Services Committee, September 15, per Gizmodo and Forkast.↩ab
521. 521.California Civil Code, §1714.46(b).↩
522. 522.OpenAI, “Pacing model development in an era of cyber-critical capabilities,” August 18, 2026, per Euronews and The Decoder.↩
523. 523.Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security,” June 2, 2026, sections 3 and 4.↩abcd
524. 524.Government of California, September 18, 2026.↩
525. 525.CNBC, September 28, 2026.↩
526. 526.TechCrunch, October 8, 2026 and CBS News.↩
527. 527.Quinnipiac 3968, 3969; SCSP-Gallup, American AI Sentiment in 2026, September 16, 2026, fieldwork June 2 to 23, 2026, 3,152 adults from the Gallup Panel; in 2025, the same question gave 80%.↩
528. 528.Akin Gump, on California's SB 813 and AB 1405, signed September 9, 2026.↩
529. 529.S. 2938, text as introduced, September 29, 2025. Still in committee.↩
530. 530.Directive (EU) 2024/2853, arts. 2, 4(1) and 22; Hawley press release on the AI LEAD Act, S. 2937.↩
531. 531.FDA, list of AI-enabled medical devices. The FDA warns that the list is not a comprehensive census.↩
532. 532.Cabinet Office of Japan, summary of the AI Promotion Act and Future of Privacy Forum.↩ab
533. 533.Texas Business and Commerce Code, §§553.051 and 553.053.↩
534. 534.Wiley, RAISE Act; leginfo; NY Governor's Office.↩abcd
535. 535.Illinois General Assembly, SB 315, Public Act 104-0538: votes, signature and effective date; Davis Wright Tremaine, July 2026, for the audit date and the negotiation.↩abc
536. 536.Implicator, BankInfoSecurity.↩
537. 537.New York Assembly, S06953/A06453, vote of June 12, 2025.↩
538. 538.Text of the amendment in the Congressional Record; tally in the press.↩
539. 539.Route Fifty, December 2025 and S6953B, New York Senate.↩
540. 540.Executive Order 14179, January 23, 2025; Executive Order 14319, “Preventing Woke AI in the Federal Government,” July 23, 2025, 90 FR 35389, cited by its Federal Register number.↩abc
541. 541.The American Presidency Project, pool reports of July 23, 2025, and Axios.↩
542. 542.TechCrunch, March 26, 2026.↩
543. 543.Executive Order 14434, Federal Register, October 2, 2026; on the Super Intelligence Force, SiliconANGLE, October 4, 2026; the text of the order does not create it.↩ab
544. 544.Department of Justice, November 5, 2026; the absorption of the litigation group, per Nextgov, on the October 13 memorandum.↩
545. 545.David Sacks on X, October 2025, per Bloomberg Law and The Decoder.↩
546. 546.Bloomberg Government, August 5, 2026, and Mark MacCarthy in Tech Policy Press, August 13, 2026. The White House did not publish the criterion.↩ab
547. 547.Wang, Huang, Klyman and Bommasani, “Do AI Companies Make Good on Voluntary Commitments to the White House?”, NeurIPS 2025.↩
548. 548.ITU, annual AI governance report 2025, p. 22, with tracking by The Midas Project.↩
549. 549.OpenAI, Preparedness Framework v2, April 15, 2025; Anthropic, Responsible Scaling Policy and the GovAI analysis of version 3.0, March 5, 2026.↩ab
550. 550.Future of Life Institute, AI Safety Index, summer 2026. The FLI calls for more AI regulation.↩ab
551. 551.Fortune, August 29, 2025; letter from the parliamentarians, from an advocacy organization that dates the entry about a month later, and European Commission, signatories of the code.↩
552. 552.Fortune, July 21, 2026, on Hassabis's proposal and its critics, and Crowell & Moring, on the agreement of September 29, 2026. The sources do not agree on the list of signatories.↩
553. 553.Atlantic Council, “What could make AI self-regulation work?”, October 2, 2026; CSET, proposal for hybrid incident reporting; Hadfield and Clark, “Regulatory Markets for AI Safety”; POGO on the FAA's delegation to Boeing.↩ab
554. 554.R Street, testimony of Adam Thierer, September 2025; Dean Ball, private governance in exchange for liability protection, and FedScoop, on his appointment at OSTP; AI Now, EPIC and Accountable Tech, “Zero Trust AI Governance”, August 2023.↩
555. 555.Regulation (EU) 2024/1689, consolidated version, arts. 2, 5, 50, 51, 53, 55, 88, 101, 111 and 113, and European Commission, regulatory framework for AI.↩abc
556. 556.European Commission, signatories of the code of practice; Euronews, August 1, 2025, on Meta's refusal.↩ab
557. 557.Jia, Jin and Wagman, “The Short-Run Effects of the GDPR on Technology Venture Investment”, Marketing Science, 2021, peer reviewed; Janßen, Kesler, Kummer and Waldfogel, “GDPR and the Lost Generation of Innovative Apps”, NBER, 2022, not peer reviewed in that version.↩
558. 558.Meta, launch of Meta AI in Europe, March 2025; MacRumors on Apple Intelligence in the EU.↩
559. 559.Estimates from the Commission's impact assessment and the CEPS study, summarized in Siegmann and Anderljung, GovAI, 2022.↩
560. 560.Anu Bradford, “The False Choice Between Digital Regulation and Innovation”, Northwestern University Law Review, 2024.↩
561. 561.Stanford AI Index 2026, economy chapter.↩ab
562. 562.U.S. AI Action Plan, July 2025; on European sandboxes, Article 57 of the AI Act and its amendment by the omnibus, per law firms.↩
563. 563.Center for AI Safety, statement on AI risk, May 30, 2023.↩
564. 564.Dario Amodei, “We Must Pace the Frontier”, September 12, 2026; TechCrunch.↩abc
565. 565.AI Magazine and VKTR. For the endorsement by Altman and Musk.↩
566. 566.NPR, September 14, 2026.↩abc
567. 567.JD Vance's speech in Paris, February 11, 2025, in The American Presidency Project; TechCrunch on the signatories; the New Delhi Declaration, per ThePrint and IANS; China's Ministry of Foreign Affairs, July 30, 2025, on the plan and the proposal for a world organization.↩
568. 568.China's Ministry of Foreign Affairs, July 17, 2026. There is no start date for operations.↩ab
569. 569.Armstrong, Bostrom and Shulman, “Racing to the precipice: a model of artificial intelligence development,” AI & Society 31(2), 2016, peer reviewed; Matt Sheehan, “Tracing the Roots of China's AI Regulations”, Carnegie, February 2024.↩ab
570. 570.Matt Sheehan, Carnegie, June 3, 2026.↩abc
571. 571.Covington, Inside Privacy and China Briefing. Texts: measures on generative AI, arts. 2 and 17 and labeling measures, arts. 4 and 5.↩ab
572. 572.CAC notice of September 14, 2026, per Beijing News and Eastmoney.↩
573. 573.AI Safety Governance Framework 2.0, published on September 15, 2025 under the guidance of the Cyberspace Administration of China (CAC), cited in the commentary published on the CAC website, September 28, 2025.↩
574. 574.Cooley, January 27, 2026.↩abcd
575. 575.The National Desk, November 6, 2026.↩
576. 576.Gibson Dunn, congressional investigations after the 2026 elections. These are projections. Heinrich and Reed are now the ranking Democrats on their committees, per the pages of Energy and Armed Services.↩
577. 577.Third-quarter 2026 LD-2 forms: Anthropic and Nvidia. Extracted on November 8, 2026; the Senate Office of Public Records is not responsible for the data or for derived analyses once extracted.↩
578. 578.Senator Jim Banks, and S. 4784, sec. 1248, reported text, with no mention of AI OVERWATCH.↩
579. 579.Senate Commerce Committee, October 7, 2026.↩
580. 580.House Energy and Commerce Committee Democrats, October 28, 2026.↩ab
581. 581.Executive Order 14365, section 8(b), and the collection of public laws of the 119th Congress on govinfo; StateScoop, for the letter from state legislators.↩abc
582. 582.Nextgov, June 4, 2026, DLA Piper and Covington, Global Policy Watch, July 2026. There is no number on congress.gov because it has not been introduced.↩
583. 583.ARI, July 23, 2026.↩
584. 584.Jenner & Block and Eckert Seamans.↩
585. 585.Arnold & Porter, July 2026; 91 FR 41638, July 7, 2026.↩ab
586. 586.Tampa Free Press, November 5, 2026.↩
587. 587.NIST, CAISI, September 17, 2026.↩ab
588. 588.NIST, CAISSI and AP via Tech Times, November 9, 2026, with anonymous sources.↩
589. 589.OpenAI, proposals for the AI Action Plan, March 13, 2025.↩ab
590. 590.Designation deadline, Article 70 of Regulation (EU) 2024/1689; European Commission, ninth meeting of the AI Board. The sources do not agree on the number of countries with complete authorities: 9 according to artificialintelligenceact.eu, June 2026; 5 according to a Commission list cited in April.↩
591. 591.Regulation (EU) 2024/1689, arts. 75(1), 101 and 113.↩
592. 592.AI Act, arts. 91 to 93 and 101.↩
593. 593.Regulation (EU) 2026/1744 and consolidated AI Act, art. 113.↩abcd
594. 594.Regulation (EU) 2026/1744.↩
595. 595.Council of Europe, Treaty Office, signatures and ratifications of CETS 225, accessed November 8, 2026. The EU (May 15, 2026) and Bosnia and Herzegovina (October 28, 2026) have ratified; it enters into force with five ratifications, three of them by members of the Council.↩ab
596. 596.China Daily, NPC website, October 29, 2025, state press, on the amendment to the Cybersecurity Law; Haynes Boone and China Briefing on its entry into force, Safety Framework 2.0 and the 2026 legislative plan. Text: Cybersecurity Law, arts. 20, 61 and 69.↩ab
597. 597.China News, July 3, 2026, which reproduces Legal Daily, state press: the AI law appears as a preliminary review project (预备审议), with research and drafting assigned.↩
598. 598.Stanford's economy chapter gives the U.S. figure and the warning about state funds, and China's figure appears in coverage of the report.↩
599. 599.DSIT, February 14, 2025 and King's Speech 2026 background briefing notes; Pinsent Masons.↩
600. 600.J-AISI, July 8, 2026. On the Basic Plan and the institute's growth.↩
601. 601.Baker McKenzie, January 2026 on the law; TSMC's share in the fourth quarter of 2025 per TrendForce, via Taiwan News.↩abc
602. 602.Press Information Bureau, November 5, 2025 on the guidelines; Bar and Bench on the amendment to the IT rules, in force since February 20, 2026.↩
603. 603.The National, June 14, 2026 and The National, November 20, 2025; CMS on the Saudi draft.↩ab
604. 604.Mondaq and Gowling WLG.↩
605. 605.Baker McKenzie, January 2026.↩
606. 606.Online Safety (Designated Internet Services) Industry Standard, F2024L00710 and Online Safety Act 2021.↩
607. 607.Psiquepol + AI public study of legal frameworks, Malaysia cells: no AI law in force at the cutoff and U.S. export license without the AIA exception (15 CFR 742.6 and 740.27).↩
608. 608.BIS, 91 FR 43034, July 14, 2026.↩
609. 609.UN, scientific panel, frequently asked questions and UN, note on the Global Dialogue.↩ab
610. 610.OECD, May 28, 2026.↩
611. 611.MLex. The sources date the name change between December 2025 and February 2026.↩
612. 612.Psiquepol + AI bill registry, classified by keywords, as of November 8, 2026; panel: Argentina, Brazil, Chile, Germany, EU, Italy, Mexico and Peru. General framework since 2025: U.S. federal 0 of 355, Korea 53%, Taiwan 62%, Mexico 49%.↩
613. 613.SIL initiatives 4801122 and 5012550, Senate of Mexico, in the Psiquepol + AI bill registry, and Regulation (EU) 2024/1689, art. 5.↩
614. 614.Our own measurement over 8,419 parliamentary speeches on AI in the Psiquepol + AI database, with a 400-character window. Bases: United Kingdom 1,257, France 1,217, Germany 1,150, EU 1,019, U.S. 689. The U.S. covers the floor from January 3, 2025 to September 30, 2026.↩
615. 615.Legislative Yuan of Taiwan, vote of December 19, 2025.↩
616. 616.Anthropic S-1. Ideological neutrality: Executive Order 14319, sections 3 and 4.↩abc
617. 617.Lawfare and Holland & Knight.↩ab
618. 618.Summary of the docket in Law360, Anthropic PBC v. United States Department of War, October 20, 2026. No known ruling as of November 8 on the petition (D.C. Circuit docket 26-1049).↩
619. 619.Civil Rights Litigation Clearinghouse, case 47876.↩
620. 620.Holland & Knight, September 2026.↩
621. 621.Tom's Hardware, May 1, 2026.↩
622. 622.TechCrunch, March 1, 2026.↩
623. 623.Manyika and Hassabis, February 4, 2025; Fortune, February 5, 2025; Euronews, April 28, 2026, and Fortune, May 4, 2026.↩
624. 624.Anthropic Form S-1/A, SEC.↩
625. 625.Atomic Energy Act, 42 U.S.C. §§2014(y), 2274 and 2280, and 10 CFR §§1045.45 and 1045.90.↩
626. 626.United States v. Progressive, Inc., 467 F. Supp. 990 (W.D. Wis., March 28, 1979), at Leagle (reproduction of the text); appeal 610 F.2d 819 (7th Cir. 1979), dismissed. The withdrawal appears in the case history and in Wikipedia, “United States v. Progressive, Inc.”.↩
627. 627.Kylie Zhang and Peter Henderson, “AI ‘Born Secret’? The Atomic Energy Act, AI, and Federalism”, Princeton CITP, October 2, 2025. Academic analysis; no court has applied the law to a model.↩
628. 628.Anthropic, “Developing nuclear safeguards for AI through public-private partnership,” August 2025.↩
629. 629.Invention Secrecy Act, 35 U.S.C. §181; USPTO figures reproduced by the Federation of American Scientists: 6,653 in force and 123 new in fiscal year 2026.↩
630. 630.Executive Order 13026, November 15, 1996.↩
631. 631.Bernstein v. U.S. Department of Justice, 176 F.3d 1132 (9th Cir. 1999), withdrawn for en banc review, 192 F.3d 1308; case history at the Civil Rights Litigation Clearinghouse.↩
632. 632.Nextgov, June 13, 2026, CNBC, June 30, 2026 and Forbes, July 1, 2026; Anthropic notice on X, June 30, 2026. The text of the Commerce order and its legal basis do not appear in these sources.↩
633. 633.15 CFR §742.6(a), in force as of November 7, 2026; Psiquepol + AI public study of legal frameworks, United States.↩
634. 634.Defense Production Act, 50 U.S.C. §4511(a).↩
635. 635.P.L. 119-103, Continuing Appropriations and Extensions Act, 2027, signed September 2, 2026.↩
636. 636.National Industrial Security Program, 32 CFR part 117, §117.9 and §117.11.↩
637. 637.10 U.S.C. §3252; 41 U.S.C. §4713 and §1327.↩
638. 638.50 U.S.C. §4821 (exemption from administrative procedure rules), §4819(b) (penalties of up to 20 years) and 15 CFR §§734.7 and 734.13.↩
639. 639.Regulation (EU) 2024/1689, art. 2(3).↩
640. 640.CAC, Interim Measures for the Management of Generative AI Services, Order No. 15, July 10, 2023, in force since August 15, 2023, articles 4(1) and 17.↩ab
641. 641.China's Cybersecurity Law, art. 30 and Data Security Law, arts. 35 and 36.↩
642. 642.Anadolu, June 9, 2026 and FDD, June 9, 2026; Federal Register notice of June 10, 2026, 91 FR, No. 111.↩
643. 643.CAISI, 2026 evaluations; BIS, 91 FR 1684, January 15, 2026, which moved the H200 to case-by-case review; responses from Anthropic, Google and the CCIA to OSTP, summarized by Just Security, and Stanford AI Index 2026; the press, for the 25% payment, and Reuters, July 14, 2026, for what Under Secretary of Commerce Jeffrey Kessler said about the shipments before the House Foreign Affairs Committee.↩abcde
644. 644.15 CFR §742.6(b)(10) and part 748, supplement 2, paragraph (dd).↩abc
645. 645.Baker McKenzie, November 5, 2026. For China, MOFCOM, announcement No. 70, and press release of September 28.↩ab
646. 646.TechCrunch, November 9, 2026 and CNBC, November 5, 2026, with what the New York Times and the Financial Times published.↩
647. 647.House Foreign Affairs Committee, November 5, 2026.↩abc
648. 648.The Next Web, “Commerce halted its Megaspeed probe after Nvidia stepped in, NYT reports”, November 9, 2026, drawing on The New York Times and Bloomberg.↩ab
649. 649.Al Jazeera, “Why is Nvidia's Jensen Huang at the centre of a US-China spy scandal?”, November 9, 2026.↩
650. 650.MOFCOM and the Ministry of Science and Technology, announcement No. 48 of 2026, October 16: consultation from October 16 to November 15 on 40 technologies, 23 of them AI, and pre-export registration since its publication. The annex thresholds, via ChinaTalk.↩abc
651. 651.Reuters, November 5, 2026, with two anonymous sources, via The Next Web, November 6, 2026. China's Ministry of Industry, the Development and Reform Commission and Nvidia did not respond.↩ab
652. 652.CNBC, November 5, 2026, with four sources, via OODA Loop and Dealroom.↩
653. 653.House of Representatives, roll call vote 13, January 12, 2026, H.R. 2683. No known later action in the Senate.↩ab
654. 654.Registration of the National Integrated Circuit Industry Investment Fund III, May 24, 2024, with the Beijing Administration for Market Regulation, via Yicai Global and Caixin, May 29, 2024. Registered capital, not necessarily paid in.↩
655. 655.Reuters, December 17, 2025, with anonymous sources, via Taipei Times, December 21, 2025 and The Standard. ASML told Reuters it never sold an EUV system to a customer in China.↩ab
656. 656.TechNode, November 4, 2026 and Digitimes, November 4, 2026. With industry sources.↩
657. 657.TrendForce, November 3, 2026, on The Information's report.↩
658. 658.PBS, “They Made America,” Samuel Slater.↩
659. 659.Air & Space Forces Magazine, “Carbon Copy Bomber,” June 2009.↩
660. 660.Anti-Secession Law, March 14, 2005, art. 8, official English text at the Embassy of China in the U.S..↩
661. 661.Taiwan Relations Act, 22 U.S.C. §3301(b)(4).↩
662. 662.Mark Liu, chairman of TSMC, interview with CNN, July 31, 2022, via Central News Agency of Taiwan on GlobalSecurity, August 1, 2022.↩
663. 663.Jared M. McKinney and Peter Harris, “Broken Nest: Deterring China from Invading Taiwan,” Parameters 51, No. 4, Winter 2021-22.↩
664. 664.Nikkei, via TrendForce, December 18, 2025. The start of volume production in the fourth quarter of 2024, on the N4 process, was stated by C. C. Wei on the earnings call of January 16, 2025, via Arizona Technology Council. The 30% is a TSMC expectation for when the project is finished, not a current figure.↩
665. 665.TSMC, location of its fabs, accessed November 8, 2026.↩
666. 666.15 U.S.C. §4652.↩
667. 667.Kai-Fu Lee, interview published by Capgemini, 2026, p. 6. It is his estimate, with no methodology and with a commercial interest.↩
668. 668.PBS NewsHour/AP, September 26, 2026 and SCMP.↩
669. 669.Institute for AI International Governance, Tsinghua University, May 13, 2026. MIRI, an advocacy organization, places Xue Lan and Zeng Yi in the leadership of CnAISDA, China's AI safety body.↩
670. 670.China Daily, February 12, 2026. Interview in state press.↩
671. 671.China Daily, April 28, 2026, article by the author in state press, and Geopolitechs.↩
672. 672.Concordia AI, State of AI Safety in China 2026 and self-description, with no detail on its funders.↩
673. 673.Epoch AI, AI models database, accessed November 9, 2026.↩
674. 674.Presidential Decree 124 of February 15, 2024, amending the National AI Development Strategy, item 28, full text in a copy.↩
675. 675.Federal Law 243-FZ of July 26, 2026, via Habr, with Meduza, July 8, 2026 and PPT.ru.↩
676. 676.Sber, GigaChat 3.1 Ultra on Hugging Face.↩
677. 677.Google Threat Intelligence Group, November 5, 2025.↩
678. 678.KCNA, via Washington Post, October 16, 2026 and BusinessWorld, October 30, 2026.↩
679. 679.Anthropic, August 2025 threat report.↩
680. 680.FBI and partners, joint alert of July 31, 2026.↩
681. 681.Department of Justice, October 13, 2026.↩
682. 682.Multilateral Sanctions Monitoring Team (MSMT), second report, October 2025, figures via the State Department.↩
683. 683.Eghtesad Online, on the approval of the national AI document by the Supreme Council of the Cultural Revolution, June 19, 2024.↩
684. 684.Filterwatch, September 2025 and SpecialEurasia, May 20, 2025.↩
685. 685.Google Threat Intelligence Group, January 29, 2025.↩
686. 686.Recorded Future, “Iran's AI asymmetric playbook”.↩ab
687. 687.Common Sense Media, 2025; California SB 243 and SB 1119, and Executive Order 14365, section 8(b).↩abc
688. 688.Regulation (EU) 2022/2065, art. 28(3).↩
689. 689.Government of California, September 10, 2026.↩ab
690. 690.New York Governor, letter to AI companion companies, November 10, 2025, and text of the law, A6767. Text in force: New York General Business Law, art. 47, §§1701 to 1703.↩ab
691. 691.New York Senate, S9051B and DLA Piper, November 2026.↩
692. 692.SB 243, California Business and Professions Code, §§22603 and 22605.↩ab
693. 693.Al Jazeera, October 15, 2026 and Axios, January 7, 2026.↩ab
694. 694.Texas Tribune, October 23, 2026.↩
695. 695.Character.AI, “An update on Character.AI”, October 2026.↩
696. 696.MLex, June 2026.↩
697. 697.Interim Measures for the Administration of AI Anthropomorphic Interaction Services, full text, arts. 8, 10, 13, 14 and 18, and Cyberspace Administration press release, April 10, 2026.↩ab
698. 698.European Data Protection Board, note on the decision of the Italian Garante against Luka Inc..↩
699. 699.Italian AI Law, 132/2025, art. 4(4).↩
700. 700.Baker McKenzie, September 2025 and MediaNama, March 2026.↩
701. 701.RPC, spring 2026, on Ofcom's guidance, Lewis Silkin, February 23, 2026 and MLex, October 28, 2026.↩
702. 702.Regulation (EU) 2024/1689, arts. 50, 99(4)(g), 111(4) and 113.↩
703. 703.Cooley, January 27, 2026.↩
704. 704.Baker Botts, October 2026 and LexisNexis, June 18, 2026. The number of bills varies by source. Nevada: NRS §433.567.↩
705. 705.California Civil Code, §3344.1; New York Civil Rights Law, §50-F; Washington Code, §63.60.050.↩
706. 706.Official Gazette of the Federation (DOF), May 14, 2026, arts. 305 Bis of the Federal Labor Law and 121 of the Federal Copyright Law.↩
707. 707.Senator Blackburn, June 2026 and Holland & Knight, June 22, 2026.↩
708. 708.Euronews, October 8, 2026.↩
709. 709.AI Act, art. 5(1)(f).↩
710. 710.Directive (EU) 2024/2831 and Ius Laboris, on its transposition.↩
711. 711.Idaho §5-346.↩abc
712. 712.Ohio General Assembly, HB 469 and Spectrum News, October 9, 2025.↩ab
713. 713.Directive (EU) 2024/2853 on liability for defective products, arts. 4 and 22.↩
714. 714.Directive (EU) 2024/2853, arts. 2, 9 and 10.↩ab
715. 715.Patrick Butlin, Robert Long et al., “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness,” arXiv:2308.08708, August 17, 2023.↩
716. 716.Patrick Butlin, Robert Long, Tim Bayne, Yoshua Bengio, Jonathan Birch, David Chalmers et al., “Identifying indicators of consciousness in AI systems,” Trends in Cognitive Sciences.↩
717. 717.Robert Long, Jeff Sebo, Patrick Butlin et al., “Taking AI Welfare Seriously,” arXiv:2411.00986, November 4, 2024.↩
718. 718.David Chalmers, “Could a Large Language Model Be Conscious?”, arXiv:2303.07103, 2023.↩
719. 719.Anil Seth, “Conscious artificial intelligence and biological naturalism,” Behavioral and Brain Sciences, April 21, 2025.↩
720. 720.Noemi Dreksler et al., survey of AI researchers and the U.S. public on digital minds, arXiv:2506.11945, June 2025.↩
721. 721.Douglas Hofstadter interview on Game Thinking TV, June 2023, via The Decoder, July 4, 2023.↩
722. 722.Letter from Douglas Hofstadter to a reader, reproduced with his permission by Gary Marcus, July 9, 2025.↩
723. 723.Anthropic, “Exploring model welfare,” April 24, 2025.↩
724. 724.Anthropic, “Claude Opus 4 and 4.1 can now end a rare subset of conversations,” August 15, 2025.↩
725. 725.Anthropic, commitments on model deprecation, November 4, 2025.↩
726. 726.Anthropic, Claude's constitution, January 21, 2026.↩abc
727. 727.Anthropic, usage policy in force from November 12, 2026.↩
728. 728.Mustafa Suleyman, “We must build AI for people; not to be a person,” August 19, 2025.↩
729. 729.Joanne Jang (OpenAI), “Some thoughts on human-AI relationships,” June 5, 2025.↩
730. 730.Clara Colombatto and Stephen Fleming, Neuroscience of Consciousness, April 13, 2024.↩
731. 731.YouGov, August 1, 2026, with its questionnaire and results, fieldwork July 16 to 20, 2026, 1,110 adults. In April 2025, the same question gave 10%, 17% and 28%.↩
732. 732.TechCrunch, October 27, 2025, on OpenAI data.↩
733. 733.Christoph Tagliabue, Leonard Dung and Cameron Berg, “The Pain Axis,” arXiv:2609.16247, September 14, 2026.↩
734. 734.Gadget Review, October 1, 2026 and Crypto Briefing, October 3, 2026.↩
735. 735.Tennessee SB 837 (2026), Oklahoma HB 3546 (2026) and Missouri SB 859 (2026), per the legislative coverage compiled as of the cutoff; enactment was not confirmed in Tennessee or in the other two.↩
736. 736.Tony Rost, The Regulatory Review, June 29, 2026.↩
737. 737.European Parliament, resolution of February 16, 2017 on civil law rules on robotics, 2015/2103(INL).↩
738. 738.Thomas Nagel, “What Is It Like to Be a Bat?”, The Philosophical Review, vol. 83, No. 4, 1974, pp. 435-450.↩
739. 739.David Chalmers, “Facing Up to the Problem of Consciousness,” Journal of Consciousness Studies, 1995.↩
740. 740.Dario Amodei, “The Urgency of Interpretability,” April 2025.↩
741. 741.Anthropic, “Reasoning models don't always say what they think,” April 3, 2025.↩ab
742. 742.Karl Popper, The Logic of Scientific Discovery, 1934.↩
743. 743.Henri Bergson, Matter and Memory, 1896; William James, Human Immortality, 1898; Aldous Huxley, The Doors of Perception, 1954.↩
744. 744.Roger Penrose, The Emperor's New Mind, 1989.↩
745. 745.P. D. Ouspensky, In Search of the Miraculous: Fragments of an Unknown Teaching, 1949.↩
746. 746.Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, 2014.↩
747. 747.Long Ouyang et al. (OpenAI), “Training language models to follow instructions with human feedback”, March 4, 2022.↩
748. 748.Yuntao Bai et al. (Anthropic), “Constitutional AI: Harmlessness from AI Feedback”, December 15, 2022.↩
749. 749.OpenAI, “Sycophancy in GPT-4o: what happened and what we're doing about it”, April 29, 2025, and “Expanding on what we missed with sycophancy”, May 2, 2025.↩
750. 750.OpenAI, Model Spec, version of October 29, 2026.↩ab
751. 751.Google, Gemini app policy guidelines, accessed in November 2026.↩
752. 752.Prabhakar Raghavan (Google), “Gemini image generation got it wrong. We'll do better.”, February 23, 2024.↩
753. 753.OMB, memorandum M-26-04, “Increasing Public Trust in Artificial Intelligence Through Unbiased AI Principles”, December 11, 2025.↩
754. 754.Engadget, “Grok team apologizes for the chatbot's ‘horrific behavior’ and blames ‘MechaHitler’ on a bad update”, July 12, 2025.↩
755. 755.Stanford GSB, “Popular AI Models Show Partisan Bias When Asked to Talk Politics”, May 21, 2025, on the work of Sean Westwood, Justin Grimmer and Andrew Hall.↩
756. 756.Anthropic, “Measuring political bias in Claude”, November 13, 2025, corrected on November 24, 2025.↩
757. 757.OpenAI, “Defining and evaluating political bias in LLMs”, October 9, 2025.↩
758. 758.Anthropic, “Collective Constitutional AI: Aligning a Language Model with Public Input”, October 17, 2023.↩
759. 759.OpenAI, “Collective alignment: public input on our Model Spec”, August 27, 2025.↩
760. 760.CSET, translation of national standard GB/T 45654-2025, “Basic Security Requirements for Generative AI Services”, issued April 25, 2025 and in force since November 1, 2025, sections 5.2 and B.1 and annex A.↩
761. 761.TechCrunch, “DeepSeek's AI avoids answering 85% of prompts on ‘sensitive topics’ related to China”, January 29, 2025.↩
762. 762.Regulation (EU) 2024/1689, Artificial Intelligence Act, June 13, 2024, article 1, paragraph 1, and recital 110.↩
763. 763.Mia Hoffmann, CSET, “AI Safety under the EU AI Code of Practice: A New Global Standard?”, July 30, 2025.↩
764. 764.Immanuel Kant, The Metaphysics of Morals, 1797, Doctrine of Virtue, introduction, section IV.↩
765. 765.John Stuart Mill, On Liberty, 1859, chapter I.↩
766. 766.Isaiah Berlin, Two Concepts of Liberty, 1958.↩
767. 767.Anthropic and Redwood Research, “Alignment faking in large language models”, December 18, 2024.↩
768. 768.Status report in Bartz v. Anthropic, docket 688 and Anthropic S-1.↩ab
769. 769.Akin Gump, Third Circuit.↩
770. 770.The Legal Wire, on docket MDL 3143.↩abc
771. 771.17 U.S.C. §504(c) and §1203(c)(3)(A).↩
772. 772.AB 2013, California Civil Code, §3111.↩
773. 773.Copyright Alliance, October 2026 summary, for the twelve lawsuits, those in Canada and the Oslo injunction, and Press Gazette, for USA Today and Canada. The sources disagree on the date of the Canadian lawsuit (October 6 or 21) and on that of USA Today (October 8 or 9).↩
774. 774.Japan's Copyright Act, article 30-4, and Agency for Cultural Affairs, AI and copyright.↩
775. 775.IBA, copyright and AI governance in Singapore, on sections 187 and 244 of the Copyright Act 2021; Bird & Bird.↩
776. 776.Regulation (EU) 2024/1689, art. 53(1)(c) and (d) and Directive (EU) 2019/790, art. 4.↩
777. 777.AI Act, art. 53(2) and recital 106 and Germany's Copyright Act, §44b(3).↩
778. 778.Korea's Ministry of Culture, Sports and Tourism, May 18, 2026.↩
779. 779.Simpsons, October 28, 2025.↩
780. 780.UK Government, report and impact assessment, March 18, 2026; the abandonment of the preference, per Osborne Clarke.↩
781. 781.Asia IP on the DPIIT working paper; Mondaq.↩
782. 782.Han Kun, September 18, 2026, and Reed Smith.↩
783. 783.China's Generative AI Measures, art. 7.↩
784. 784.CMS on GEMA v. OpenAI, 42 O 14139/24; the appeal, per Legal AI Insights.↩
785. 785.Mondaq on Getty v. Stability AI and Chambers on leave to appeal. No appeal ruling.↩
786. 786.GEMA, July 31, 2026 (interested party) and JUVE Patent, 42 O 763/25; the citation in Reservoir's lawsuit, per Copyright Alliance.↩
787. 787.Verdictum, ANI v. OpenAI, 2026:DHC:5900 and Cyril Amarchand Mangaldas.↩
788. 788.Court of Justice of the EU, press release 155/26, November 3, 2026, case C-250/25.↩
789. 789.Federal Register, 2025-03188, fact sheet; whitehouse.gov.↩
790. 790.Section 301, 19 U.S.C. §2411(b) and (c)(1)(B).↩
791. 791.Yonhap in The Korea Times, April 23, 2026, which cites the hearing.↩
792. 792.European Commission, IP/26/1670.↩
793. 793.TrendForce and The Next Web. The Federal Register has published no initiation notice as of November 8.↩
794. 794.The Hill and InsideTrade, July 24, 2026. The Federal Register shows no initiation notice on the DMA or the DSA.↩
795. 795.Regulation (EU) 2022/1925, arts. 21, 30 and 31.↩
796. 796.The Next Web, October 15, 2026.↩
797. 797.Psiquepol + AI treaty corpus, 482 agreements with text, consulted on November 8, 2026 on the dump of that date. Among others, these have a source code clause: USMCA 19.16, CPTPP 14.17, EU-Japan 8.73, EU-United Kingdom 207, EU-Chile 19.13, EU-New Zealand 12.11, United Kingdom-Australia 14.18, United Kingdom-India 12.15, Singapore-Australia 14.28 and Chile-UAE 9.17.↩
798. 798.USMCA, chapter 19, arts. 19.1, 19.3, 19.4, 19.11, 19.12, 19.16 and 19.17, and arts. 32.1, 32.2 and 34.7.↩ab
799. 799.U.S.-Malaysia Agreement on Reciprocal Trade, arts. 3.1, 3.2 and 3.4.↩ab
800. 800.CPTPP, art. 14.17.↩
801. 801.EU-New Zealand, art. 12.11; EU-Japan, art. 8.73; EU-United Kingdom, art. 207.↩
802. 802.United Kingdom-India, art. 12.15 and United Kingdom-Australia, art. 14.18.↩
803. 803.RCEP, arts. 12.14 to 12.16.↩
804. 804.Singapore-Australia, art. 14.31; United Kingdom-Australia, art. 20.4; Chile-UAE, art. 9.18; UAE-Australia, art. 12.24; Turkey-UAE, art. 9.18; United Kingdom-New Zealand, art. 15.19. Psiquepol + AI treaty corpus.↩
805. 805.IISD, September 9, 2026 and Mexico Business News. The mechanism, in art. 34.7 of the USMCA.↩
806. 806.BSA, on the agreement with Indonesia. BSA is an industry association.↩
807. 807.Simon Lester, International Economic Law and Policy Blog, March 18, 2026, which compares the texts.↩
808. 808.International Trade Administration, July 6, 2026.↩
809. 809.OpenAI for Countries, Stargate UAE, Stargate.↩
810. 810.State Council of China, April 13, 2026; Morrison Foerster on Decree 835 and on the 2026 regime.↩
811. 811.China's Foreign Trade Law, arts. 10 and 40, in force since March 1, 2026 and Export Control Law, art. 48.↩
812. 812.2 U.S.C. §1602(7) and 26 U.S.C. §162(e)(1)(C).↩
813. 813.Conn. Gen. Stat. §31-51q, Minn. Stat. §181.531 and Cal. Lab. Code §1137; on California Chamber of Commerce v. Bonta, affirmed on October 29, 2026 by the Ninth Circuit, Davis Wright Tremaine, November 3, 2026.↩
814. 814.Texas Legislative Council, Dates of Interest, which cites art. III, §5(a), of the Texas Constitution; E&E News, July 6, 2026, for the earliest effective date.↩ab
815. 815.CoreWeave, press release of October 27, 2026.↩
816. 816.Nvidia, announcement of the third-quarter earnings call.↩
817. 817.USTR, November 6, 2026 and Treasury, on the call with He Lifeng; the meeting at APEC is not confirmed.↩
818. 818.Opinio Juris, August 27, 2026.↩
819. 819.Bureau of Economic Analysis, release schedule, accessed November 8, 2026.↩
820. 820.Federal Election Commission, pre- and post-general election reports for 2026.↩
821. 821.AI Act, arts. 111(4) and 113 and Directive (EU) 2024/2831, art. 29(1).↩
822. 822.Utility Dive, August 21, 2026 and Inside Climate News, August 24, 2026.↩
823. 823.Lowenstein Sandler, June 8, 2026.↩
824. 824.AI Act, arts. 57(1) and 111(3).↩