Recently, I watched a YouTube video in which Fields Medallist Hugo Duminil-Copin presented A Severe Misalignment of AI in Mathematics, listened to Gina Raimondo on Freakonomics and read Brett Hemenway Falk and Gerry Tsoukalas's The AI Layoff Trap. These were different discussions, but together they left me wondering about the same thing: are we preparing quickly enough for the changes ahead? The question does not just pertain to AI’s capability. Rather it is what that capability might mean for people whose income, training and sense of purpose depend on work that machines will increasingly perform.

If AI allows businesses to produce more software, information and discoveries with much less human labour, people could work fewer hours and have more time for their families, learning and other activities. But there are other possible paths. If machines do much of the work, the people who own those machines may receive most of the money they generate. Those owners might be only a small group whose interests do not reflect those of society as a whole. Meanwhile, people who depend on their salaries could find themselves earning less or struggling to find work at all. This could eventually create problems even for the owners of the machines. Businesses do need customers who can afford to buy what they produce.

Thus, two connected problems should be addressed as a matter of some urgency. First, how do we help people who lose their jobs, earn less or struggle to find work because of AI and AI-related systems? Second, as machines do more and more of the work, who will receive the wealth they create? Helping people find another job is hardly acceptable if there are fewer jobs available, and protecting their income today does not answer the question of who will own tomorrow's technology and what happens to society tomorrow.

These two challenges can be dealt with together. First, Europe should create a system to protect people's incomes when AI disrupts their working lives. This could help in the following circumstances: loss of a job, working hours reduced or absence of employment opportunities. Governments would pay for this protection from the beginning, rather than waiting for AI investments to become profitable. Second, Europe should create a public investment fund that would invest in AI companies and the infrastructure they need, such as powerful computers, chips and semiconductor technology, data centres, robotics, and energy supplies. The fund would hold these investments for the benefit of European citizens. Over time, returns on the public investments could supplement the protection system. The aim is to prepare for disruption while giving Europe a stake in the technology shaping its future.

To understand why there is some urgency in addressing these problems, it is worth looking at a profession that few people would have expected AI to transform so quickly: mathematics. Professional mathematicians form a small and highly specialised community, but what is happening to them may give us an early glimpse of changes that could eventually affect many other professions. Most of us remember mathematics from school as an abstract subject involving series of exercises. The teacher explains a method, gives us a problem and expects us to find the answer using the method. We may struggle with calculation, but we know that someone has already solved the problem. Professional mathematicians work in a very different world. They investigate questions for which nobody knows the answer, sometimes questions that have remained unanswered for decades or even centuries.

These questions often come in the shape of a pattern, involving numbers or shapes, that appears multiple times. Mathematicians test more examples and begin to form the idea that the pattern might always hold. This idea is called a “conjecture”: a statement that seems to be true but has not yet been proved. We do not yet know whether the pattern will always hold. Conjectures are important because, for example, they may reveal a hidden connection between different areas of mathematics or open the door to discoveries that nobody had previously imagined. Even a conjecture that turns out to be wrong can teach mathematicians something important.

To establish that the conjecture is true, the mathematician must develop a “proof”: a chain of reasoning showing why the result must hold in every case covered by the statement, not just that it works a million times. Alternatively, the mathematician might discover a single example where the rule fails. This is called a “counterexample”, and just one valid counterexample is enough to show that a conjecture is false. Finding either a proof or a counterexample can be extraordinarily difficult with some conjectures finding no proof or counterexamples for decades. Some mathematicians spend years on conjectures. This involves testing ideas, searching for connections, developing new methods or approaches. That is why beyond finding answers, mathematical research is also about developing ideas, learning new ways of thinking and understanding why something is true or false.

This helps explain why recent AI developments caused simultaneously much excitement and anxiety. On 20 May 2026, OpenAI announced that one of its models had disproved the Erdős unit-distance conjecture, a long-standing claim about how many pairs of points in a plane can be exactly one unit apart (OpenAI, 2026a). OpenAI described an unexpected construction, and external mathematicians checked the proof and wrote a companion paper about its significance. In it they describe a surprising connection made by AI to algebraic number theory. Tim Gowers called the result a milestone and said he would have recommended acceptance had a human submitted it to the Annals of Mathematics (Alon et al., 2026). Could this be closer to the kind of creative reasoning that mathematicians spend years developing? In August, OpenAI said the May disproof had already prompted further mathematical work and reported ten additional results (OpenAI, 2026c). By 6 October, OpenAI had released 719 manuscripts organised into 372 related families of results, produced with an internal model (OpenAI, 2026b). The scale raises a new question: how can mathematicians check, understand and make use of such a large body of work?

The Mathematics Example As A Warning Sign

AI is beginning to participate in the creative stages of research, not simply accelerate routine mathematical work. Systems can search literature, test approaches and connect ideas; the May proof suggests they can also uncover routes that surprise specialists (OpenAI, 2026a). The mathematics example may also point to something more radical than ordinary mechanisation. If we look back at the introduction of tractors, that new technology replaced much manual labour, but a tractor still needed a driver and a farmer deciding what to grow and how to use the crop. If AI increasingly moves from carrying out intellectual tasks to choosing problems, producing solutions and eventually using those solutions itself, the analogy begins to look more like a tractor that does not just replace field help, but learns to drive itself and then to run the farm. This is a possibility that rapid iterations in advances in robotics bring closer to reality every month. It does explain why the discussion reaches beyond whether machines can complete an existing task faster, towards who decides what work to undertake and how to use its results.

The concern did not begin with the October release. On 11 September, an initial group of 25 Fields Medallists endorsed A Severe Misalignment of AI in Mathematics, drawing attention to the difference between showing what a model can do and supporting a mathematical community (Math and AI, 2026). A company can showcase its system through solved problems, while mathematicians may value the ideas, training and further questions that emerge from working on them. What strikes me here is that delivering an answer may not serve both purposes. Indeed, receiving a finished piece of music does not teach someone to compose, and a mathematical problem can likewise be both a question awaiting an answer and a path to becoming mathematicians. Mathematicians and composers experiment, fail, develop techniques and might even discover something new. In that sense the process of trying to solve an unsolved problem has value independently of whether it eventually produces the answer.

This tension also highlights a related but distinct concern; deskilling. This concern generally applies to how humans now interact with AI. If mathematicians increasingly use or merely verify proofs produced by AI rather than find proofs themselves, some of the skills involved in discovery may gradually weaken. This is similar to modern car diagnostic systems that can identify a faulty component and tell a mechanic what to replace. That makes repair faster, but over time a mechanic who follows diagnostic instructions may understand less deeply why an engine is behaving as it does than an earlier mechanic who had to diagnose it from sound, behaviour and mechanical principles. These concerns do not justify hiding answers to mathematics problems, but they explain why a disagreement about publication can become a disagreement about what a profession is for. Producing the result and helping people learn through the process are connected but they are not the same thing.

Questions about the publication process also concern who controls it. On 29 September, the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) called for responsible publication, proper credit for others' work and funding for human understanding. It also opposed testing advanced problems on company-owned models that other researchers could not access (AGMAI, 2026a). After the October release, it described publication as the beginning of understanding rather than its completion, and argued that mathematicians should still choose their own questions (AGMAI, 2026b). The Association for Human Mathematics viewed publishing hundreds of documents at once constituted a display of corporate power rather than satisfactory scholarship (Association for Human Mathematics, 2026). Its argument is that if a company can change a profession, people working within it need some influence over the constituents of that change.

The reactions within research communities made those concerns less abstract. Thomas Bloom reported that his Erdős-problems website increasingly hosted generated proofs rather than discussion, so he paused new comments and proof claims and announced changes intended to restore participation (Bloom, 2026). A good dish may be perfectly prepared, yet whether someone wants to eat it can still depend on how it is served. Mathematical results may be similar: the result itself matters, but the way it is presented can influence whether others want to engage with it, understand it and build on it. Álvaro Lozano-Robledo reported students reconsidering mathematical careers alongside excitement about new opportunities (Lozano-Robledo, 2026). Those responses can coexist. Someone can welcome a discovery while worrying about students trying to enter the profession. None of this establishes that mathematicians have already lost their job or income. But it does raise questions about investments, grants or doctoral projects. The employment question begins there, in the decisions about whose work receives support and how someone starting out can find a place, rather than only in the number of researchers who have already received dismissal notices.

Universities pay salaries, agencies award grants and doctoral programmes support students, partly on the expectation that people will spend months or years tackling difficult problems. Suppose a machine completes much of that work in hours. A mathematician may keep today's contract while the case for the next one changes; a researcher may need another programme, a student may find that someone has solved the thesis question, and a funder may reconsider how much human time a project needs. Useful human activity does not guarantee paid opportunities to perform it. The same possibility arises outside mathematics: if one lawyer using AI can handle work that previously required ten, nine people need not receive redundancy notices because their jobs may never open. Software firms and laboratories could likewise reduce recruitment before dismissing existing staff. Indeed, if much of the remaining human work involves checking, interpreting or validating AI outputs, employers will retain experienced employees while hiring fewer junior ones. These are illustrations, not estimates of jobs already lost, but they show why watching layoffs alone can miss changes in the routes through which people begin earning a living.

How Can History Shed Light?

History teaches us that earlier technologies eliminated jobs but eventually created others. The difficulty is living through the “eventually”. But nothing guarantees that history will repeat itself in the case of AI. Whether other jobs appear depends partly on there being new tasks for people to perform. Acemoglu and Restrepo (2019) explain that while automation replaces some human tasks, the creation of new ones can restore demand for workers. If AI and robotics become capable across most cognitive and physical tasks, that process may become weaker. In the Netherlands, agricultural employment fell as farms mechanised, while expanding industries offered alternative jobs (CBS, 2017; Van Meeteren and Smit, 2025). Northern Sweden offers another example: its growing timber industry provided rural workers with paid employment in logging and sawmills (Sundvall, 2024). In both cases, other industries were expanding. We cannot assume AI will create enough accessible new work in the same way.

Robert Allen describes a period during Britain's Industrial Revolution when output per worker rose but real wages stagnated (Allen, 2009). Greater productivity did not immediately translate into higher living standards. The Swing Riots show a different danger: threshing machines removed winter work from rural labourers, and unrest followed in 1830 and 1831. Poverty and low wages also contributed, but Caprettini and Voth (2020) find that the spread of these machines increased riot risk, especially where alternative employment was scarce. The first example concerns who shares in growing wealth; the second concerns what happens when people cannot find other work.

Industrialisation also changed who owned productive assets and who depended on wages (Hobsbawm, 1975). These strains contributed to wider debates about ownership and workers' rights, although the revolutions of 1848 had many political and economic causes (Hobsbawm, 1962). Labour movements and public institutions gradually helped provide answers: collective bargaining, education, unemployment insurance, pensions and healthcare offered protection that machines did not provide on their own (Marshall, 1950; Esping-Andersen, 1990).

This returns us to Acemoglu and Restrepo's (2019) distinction between jobs displaced by automation and new tasks that create opportunities for people. Employment and wages depend partly on the balance between them, but workers losing one kind of job must also be able to reach the new opportunities. Different activities present different obstacles to automation, so the mathematics example remains an early warning rather than a labour-market forecast. In the same vein, the International Labour Organization (ILO) estimates that one in four workers worldwide has some occupational exposure to generative AI, with 3.3 per cent of global employment in its highest exposure category (Gmyrek et al., 2025). Exposure concerns tasks that AI could perform; it does not predict the share of jobs that will disappear. Because most occupations still require human input, the researchers expect transformation more often than complete replacement.

What troubles me is the possibility that opportunities narrow before anyone receives a dismissal notice. Employers can leave vacancies unfilled, recruit fewer graduates, cut paid hours or commission less freelance work, and those changes matter alongside unemployment. US payroll evidence points to one such narrowing entry route. Through June 2026, employment among 22- to 25-year-olds in AI-exposed occupations stood 19 per cent below where it would have been had it kept pace with less-exposed peers, mainly because of reduced hiring rather than increased separations (Brynjolfsson et al., 2026). Experienced workers showed no comparable gap. The authors found no widespread displacement and described the results as descriptive, not causal: they show a pattern without proving that AI caused it. Other evidence complicates any claim that the same pattern is already appearing everywhere. Danish research found no significant average change in earnings or recorded hours two years after ChatGPT's launch, ruling out effects larger than 2 per cent, even though employers were reorganising tasks (Humlum and Vestergaard, 2025). The Organisation for Economic Co-operation and Development (OECD) also notes that young graduates' rising unemployment risk predates generative AI (OECD, 2026), so the technology cannot account for the whole trend. These studies do not establish an economy-wide collapse, but neither do they make recruitment, changing tasks or entry opportunities irrelevant. They show why we need to follow how employment develops rather than wait for one decisive unemployment figure. The practical question is how to prepare for uneven changes.

If employment can shrink through recruitment, why not ask businesses to slow down? Would it not be rational for a company to ensure that people have jobs and income to continue to pay for the company’s products? It is rational but is not as easy to implement as it sounds. An individual firm may have little room to do so. In The AI Layoff Trap, Brett Hemenway Falk and Gerry Tsoukalas examine what happens when companies replace workers with AI. A company may save money by employing fewer people, but those who do not work or lose their jobs lose income and have less money to spend. Other businesses then lose customers and may also cut jobs. Eventually, even the companies that first benefited from AI could suffer because fewer people can afford to buy their products (Hemenway Falk and Tsoukalas, 2026a). Imagine two companies making similar goods. One keeps workers whom AI could replace; the other automates, lowers costs and gains room to cut prices, invest or survive on smaller margins. The first company's restraint may leave it less competitive even if both recognise the wider danger. National borders exacerbate this pressure. A firm rejecting a productive technology may lose business to many more rivals abroad that operate under different regulatory frameworks. Thus, asking or even imposing on employers to act responsibly does not, by itself, remove the difference between what benefits one company and what benefits society.

The authors consider a Pigouvian tax as a solution. This would be a corrective automation tax to make employers account for the loss of spending their decisions impose on other firms (Hemenway Falk and Tsoukalas, 2026a). Applying it, would require officials to distinguish replacement from useful assistance, consider jobs never advertised and coordinate across countries. This may prove difficult if not impossible. International Monetary Fund (IMF) researchers warn against a blanket AI tax, favouring stronger protection, taxation of capital income and a review of tax advantages that encourage displacement (Dabla-Norris and de Mooij, 2024). Further, the model itself examines one sector with similar firms rather than the whole economy. When a company spends less on wages, that money may be used elsewhere. It can become profit, investment or lower prices, and this may create other jobs. But this does not necessarily happen quickly enough to help the workers who have just lost their income.

Protecting workers and changing companies' decisions are connected but different policy tasks are required. For example, a fixed universal basic income would not, by itself, remove a company's incentive to automate in the model, but that does not make it useless against poverty. The authors also examine basic income funded from profits in a scenario where AI replaces most human labour (Hemenway Falk and Tsoukalas, 2026a, sections 4.2 and 4.7). Income security would still leave questions about professional status, expertise and opportunities to contribute. A job gives people more than money: it can also give them a role, skills, purpose and a place in society. Korinek and Stiglitz examine how technological gains and changing rewards to labour and ownership can deepen inequality (Korinek and Stiglitz, 2017). We should safeguard people because their lives matter, not only because businesses need their spending. The case for protection should not depend on proving that hardship eventually becomes unprofitable for those who first benefit from it.

These limits make public action necessary, but its timing presents another difficulty. Businesses risk losing customers if they reject productive technologies, whereas governments need to debate, consult, investigate, legislate and assign budgets. Yet waiting years for employment figures or a consensus can leave people without support after career choices and investment have already shifted. Where the risk is identified, transition tools could be an appropriate anticipatory answer. In the US, such a tool, RAISE US, was presented in the Freakconomics discussion with Gina Raimondo. RAISE US is a nonprofit organisation she and Eric Holcomb launched on 25 June 2026 (RAISE US, 2026). It brings employers, governments, educators and researchers together, linking training to employer demand, encouraging firms to retain and redeploy workers, and proposing wage insurance when new work pays less. At launch, it reported multi-year commitments exceeding half its $1 billion target. Its appeal is that it seeks to organise a response before individual workers face the consequences alone, while linking training to openings employers can actually identify.

My concern with retraining is the gap between acquiring a skill and finding paid work. If 100 people compete for 20 jobs, training may change who obtains them, but it does not produce the other 80 jobs, and a new skill offers no guarantee of lasting protection, especially in a fast-changing labour market landscape. RAISE US can test responses, but it cannot establish a universal right in law or require every company to contribute. Its safeguards include labour representation through Liz Shuler, president of the American Federation of Labor and Congress of Industrial Organizations (AFL-CIO), and a commitment that corporate contributions will not finance its Policy Lab (RAISE US, 2026). These arrangements matter without providing the certainty of a well-funded public entitlement. Philanthropy can support experiments and employers can help locate vacancies. Governments, however, still have to decide what people can claim as a right.

Europe already has experience organising support while disruption unfolds. During the Covid19 pandemic, Support to mitigate Unemployment Risks in an Emergency (SURE) lent Member States €98.4 billion to support employment and income. The programme ended in December 2022 and shows what joint action can achieve. AI is not Covid19. The pandemic and the restrictions it brought interrupted jobs that governments expected to resume. AI, on the other hand, may permanently reduce demand for some activities. A new instrument cannot simply keep every employment relationship intact until a normality returns. What matters to me about SURE is its timing, not its replication: support can begin before a shock runs its course. A European proposal could apply that again by starting income protection and public AI investment simultaneously, through linked but separate institutions. The protection arm would guarantee support while the investment arm builds assets and helps finance it, without requiring governments to identify every future loss before establishing the means to respond.

The protection arm would maintain an income floor even when a job disappears, including for people entering work rather than only those with long contribution histories. Alongside that floor, temporary earnings insurance could soften an abrupt loss. If, for example, someone previously earned €3,000 a month and found work paying €2,200, a scheme could offset part of the €800 difference for a specified period. The aim would be time to adjust, not an indefinite guarantee of every previous salary.

Support also matters during retraining because learning does not suspend expenses. In fact, IMF researchers argue for stronger, more portable protection during potentially prolonged transitions (Dabla-Norris and de Mooij, 2024). Workers could also benefit from AI by working less, rather than only by earning more. For example, if AI helps a team finish in 30 hours the work that used to take 40, the company could choose to let people work fewer hours instead of cutting jobs. But this only works if AI has genuinely saved time and money. And we cannot assume employers will voluntarily share time savings with workers. Governments could support negotiated reductions in working time instead of redundancies. Any subsidy or tax relief should, however, depend on evidence. A company should have to show when the AI system was introduced, what work it changed, what staffing and working hours looked like beforehand, and what happened afterwards. Payroll and working-time records, together with employee or works-council verification where available, could be used to check that the productivity gain is real and that public support is preserving employment rather than subsidising cuts that would have happened anyway. Payments could be withdrawn or reclaimed where the conditions are breached.

Sharing productivity gains within companies is only one part of the response. Social protection may also need new sources of funding if income shifts away from wages.

Income Protection and Shared Ownership

Paying for such protection cannot depend entirely on wages if income increasingly flows to owners. Governments would need to reconsider taxes on employment, profits and other ownership income, while making commitments against realistic revenues rather than hoped-for AI riches (Dabla-Norris and de Mooij, 2024). But this proposal is not simply about how to pay benefits. Hobbes used the idea of a social contract to explain why people might accept a common political authority in exchange for peace and security (Lloyd and Sreedhar, 2022). AI raises a different but related question: if technological progress changes how wealth is created and who receives it, what protection, voice and opportunity should citizens expect? If society shares the risks, it should also share in the gains. The important distinction lies between producing enough for everyone and giving everyone a claim on the wealth that production generates. An automated economy could achieve the first while failing at the second. One way to put that principle into practice would be a European AI Sovereign Wealth Fund, inspired by Norway's sovereign wealth fund. Norway invests petroleum revenues for present and future generations; Europe could similarly build lasting public assets in AI. The aim is public investment alongside private enterprise, not nationalisation.

Given the investment needed, one preliminary option would be an 80% holding for a public EU investment entity, with up to 10% reserved for voluntary subscriptions by EU citizens and a further 10% for Member States individually. These percentages are not a proven optimum: they would need financial and legal testing. The purpose of the citizen allocation is straightforward—to raise additional capital from Europeans willing to invest, rather than rely on non-EU investors, without allowing private shareholders to steer the fund. The public holding would remain under democratic oversight for the benefit of everyone, including people who cannot afford shares. A citizen subscriber, by contrast, would be making a private investment and could gain or lose money; dividends would not be guaranteed. The citizen allocation would be capped, as could individual subscriptions, and any voting rights would be restricted to financial matters rather than the fund's public mandate. Initial capital could also come from the EU budget and voluntary Member State contributions. The European Investment Bank could provide loans or project financing without necessarily becoming an owner, while private investors could participate in individual projects without controlling the fund. EU institutions would set the public mandate, with professional managers responsible for investments, drawing on Norway's separation of oversight and management (Norges Bank Investment Management, n.d.). The final capital structure and investor rights would need approval by the relevant EU institutions and participating Member States.

Initially and as soon as possible, the fund would invest in the infrastructure and businesses needed to establish greater European independence in artificial intelligence. This could include advanced computing facilities, European AI models, cloud infrastructure, chip and semiconductor technologies, energy generation and stakes in promising European companies. The fund would not merely distribute subsidies. It would seek to own productive assets capable of generating lasting economic value. European businesses, particularly smaller companies, could gain preferential access to computing power and technical resources that would otherwise be beyond their reach. In exchange, recipient companies could agree to proportionate, measurable commitments towards workers. These might include sharing documented productivity gains, maintaining agreed employment levels, or financing retraining and redeployment. Firms would report staffing, hours, pay and relevant output before and after receiving support; independent checks could verify compliance, with preferential access withdrawn or financial benefits recovered where commitments were breached.

As the fund's investments mature, income could come from dividends, equity gains, computing services and other commercial activities. This fund would also finance strategic European infrastructure. Its commercial and public-policy objectives would need to be assessed separately. Following Norway's example, a clear spending rule could determine how much of the fund's long-term returns could be transferred to public budgets, while preserving its capital for future generations. Those transfers could contribute to the income protection, training and employment programmes described earlier in this article. Returns attributable to privately held shares would, of course, belong to their investors, while revenues from the public holding would serve the collective interest. Public authorities would continue to finance immediate social protection through ordinary budgets, rather than make people's livelihoods depend on uncertain investment performance.

A Social Contract People Can Shape

Beyond financing and investment, the proposal calls for a public agreement on how AI is developed, how businesses benefiting from public support treat workers, and how people retain opportunities to contribute. The fund is one tool within that agreement. It is not a sovereign, and AI itself would not decide people's rights. Elected representatives would set the fund’s mandate; independent professionals would manage investments under public rules, with citizens able to see and challenge those decisions.

The transition would not be successful if people could pay their bills but found no place to contribute. Amartya Sen's work helps explain why income is a means to pursue opportunities, not an end in itself: what matters also is what people are genuinely able to do and become (Sen, 1999). My brother used to teach yoga and tai chi in villages in France. Given the choice, that was the work he wanted to do. He would teach little groups in community halls once or twice a week. He had keen students, but there was still too little demand for him to earn a living from. He had to change occupation. His experience shows how meaningful activity and sufficient paying demand may not necessarily coincide. Income protection could give someone in his position more room to choose, but it would not fill classes or guarantee a salary for every preferred or vocational occupation.

That question extends to public services and opportunities to participate. Care, environmental restoration, education, cultural work and community projects offer ways to contribute that a transition focused only on purchasing power and free-market forces often overlook. What we value and how we contribute to society may not be set by market demand forces but by our own inclinations. Could we move to a society where the market becomes less meaningful, life not reduced to performing an activity to have the means to live, and individual aspirations more easily fulfilled? Mathematics brings the same distinction back into focus. Indeed, people still need to explain ideas, teach students, check claims and decide which new questions deserve attention. A growing collection of results cannot replace communities capable of understanding them. Funding should let researchers redirect projects that AI has changed, while access to tools should not depend entirely on the goodwill of the companies controlling them. AGMAI argues that human understanding should be supported through independent institutions rather than allowing privately held AI laboratories or companies to direct that work (AGMAI, 2026a).

The credibility and acceptability of the scheme would depend on who makes decisions and who takes responsibility for such decisions. Governments and European Union (EU) institutions would agree legal powers, contributions, distributions and cross-border responsibilities before launch, rather than assuming benefits should follow the location of an investment. Elected lawmakers would decide what people are entitled to receive and set budgets, while independent investment managers answer to public oversight. Workers, unions, civil society and affected citizens would help set priorities through direct voting mechanisms and petitions. They would also have an effective ability to scrutinise decisions. Employers would report adoption and recruitment changes. A team of recognized experts in their own fields would study the technical, economic and social effects, and independent reviewers would check both the books and decisions. Fully published results plainly presented and written in plain language would let people challenge both investment choices and failures in protection.

The scheme would also need to be adaptive, evolving as the employment situation changed. Data on hiring, earnings, paid hours, graduate recruitment, research grants and entry-level opportunities could help show whether people were gaining access to paid work or losing it, because unemployment alone is too narrow a measure. Instead of waiting for an established job to disappear, such evidence could guide changes to protection, investment, shorter-hours agreements and earnings insurance. Further, Member States could learn from one another's experience, while joint financing could help countries with fewer resources provide dependable protection and reduce the concern that acting alone would drive investment or taxable income abroad. The mathematics releases do not tell us how many jobs AI will replace, nor do they prove that mass unemployment is approaching. They do show how quickly activities that once required years of human expertise may change. We do not need to predict an employment collapse to recognise the need to prepare. If the future of work is uncertain, that is a reason to build institutions that can protect people, rather than wait until the consequences are impossible to ignore.

These discussions eventually reached the family table. The youngest member offered a simpler response: “Why don't we just turn it off? Ban AI.” She said she did not need it. If AI is merely useful, something that is good to have, why accept the risks it poses to society and to people? Her question deserves an answer. She does not make these decisions, but she will live with their consequences. Her generation should help shape what comes next. The next breakthrough should give her generation a stake in its rewards, not another reason to fear next month's rent.

References

Archives municipales de Lyon (n.d.) « 1831–1834, la révolte des Canuts ». Available at: https://www.archives-lyon.fr/pages/1831-1834-la-revolte-des-canuts.

Acemoglu, D. and Restrepo, P. (2019) ‘Automation and new tasks: How technology displaces and reinstates labor’, Journal of Economic Perspectives, 33(2), pp. 3–30.

AGMAI (Advisory Group on Mathematics and Artificial Intelligence) (2026a) Responsible release of AI-generated mathematics, 29 September. https://agmai.org/general-sep29/ (accessed 13 October 2026).

AGMAI (Advisory Group on Mathematics and Artificial Intelligence) (2026b) On OpenAI's release of mathematical results, 6 October. https://agmai.org/ (accessed 13 October 2026).

Allen, R.C. (2009) ‘Engels’ pause: Technical change, capital accumulation, and inequality in the British Industrial Revolution’, Explorations in Economic History, 46(4), pp. 418–435. https://doi.org/10.1016/j.eeh.2009.04.004.

Alon, N., Bloom, T., Gowers, W.T., Litt, D., Sawin, W., Shankar, A., Tsimerman, J., Wang, V. and Matchett Wood, M. (2026) ‘Remarks on the disproof of the unit distance conjecture’. arXiv:2605.20695. https://arxiv.org/abs/2605.20695.

Association for Human Mathematics (2026) ‘AHM statement on OpenAI’s October 6 release of mathematical documents’. Statement reposted 7 October 2026: https://terrytao.wordpress.com/2026/10/07/ahm-statement-on-openais-october-6-release-of-mathematical-documents/.

Bloom, T. (2026) ‘Changes to the Erdős problems web site’, What’s New, 6 October. https://terrytao.wordpress.com/2026/10/06/changes-to-the-erdos-problems-web-site/.

Brynjolfsson, E., Chandar, B. and Chen, R. (2026) Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab working paper, revised 12 August 2026; originally published 2025. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/.

Caprettini, B. and Voth, H.-J. (2020) ‘Rage against the machines: Labour-saving technology and unrest in industrializing England’, American Economic Review: Insights 2(3), pp. 305–320. https://doi.org/10.1257/aeri.20190385.

Dabla-Norris, E. and de Mooij, R. (2024) ‘Fiscal policy can help broaden the gains of AI to humanity’, IMF Blog, 17 June. https://www.imf.org/en/Blogs/Articles/2024/06/17/fiscal-policy-can-help-broaden-the-gains-of-ai-to-humanity.

Esping-Andersen, G. (1990) The Three Worlds of Welfare Capitalism. Princeton, NJ: Princeton University Press.

European Commission (n.d.) SURE: European instrument for temporary Support to mitigate Unemployment Risks in an Emergency. https://economy-finance.ec.europa.eu/eu-financial-assistance/sure_en (accessed 13 October 2026).

Gmyrek, P., Berg, J., Kaminski, K., Konopczynski, F., Ladna, A., Nafradi, B., Roslaniec, K. and Troszynski, M. (2025) Generative AI and jobs: A refined global index of occupational exposure. ILO Working Paper 140. Geneva: International Labour Organization. https://doi.org/10.54394/HETP0387.

Hales, T. (2026) ‘What mathematicians should know about the Lean Theorem Prover: Questions of reliability and AI’, What’s New, 9 October. https://terrytao.wordpress.com/2026/10/09/what-mathematicians-should-know-about-the-lean-theorem-proverquestions-of-reliability-and-ai/.

Hemenway Falk, B. and Tsoukalas, G. (2026a) The AI layoff trap. Working paper, version 3, revised 3 June. arXiv:2603.20617v3. https://arxiv.org/abs/2603.20617v3.

Hemenway Falk, B. and Tsoukalas, G. (2026b) ‘A note from the authors of “The AI layoff trap”’, Ciphermind, 14 April. https://ciphermind.org/blog/ai-layoff-trap/.

Hobsbawm, E.J. (1962) The Age of Revolution: Europe 1789–1848. London: Weidenfeld and Nicolson.

Hobsbawm, E.J. (1975) The Age of Capital: 1848–1875. London: Weidenfeld and Nicolson.

Humlum, A. and Vestergaard, E. (2025) Still waters, rapid currents: Early labor market transformation under generative AI. NBER Working Paper No. 33777, current version consulted 13 October 2026. https://www.nber.org/papers/w33777.

INSEE (2026) ‘Professions et secteurs d’activité’, Emploi, chômage, revenus du travail, 2 July. https://www.insee.fr/fr/statistiques/8733043?sommaire=8733125.

Jarrige, F. (2013) ‘Briseurs de machines au XIXe siècle’, La Revue du projet, no. 25, March.

Korinek, A. and Stiglitz, J.E. (2017) Artificial intelligence and its implications for income distribution and unemployment. NBER Working Paper No. 24174.

Lozano-Robledo, Á. (2026) ‘What should we tell our students?’, What’s New, 8 October. https://terrytao.wordpress.com/2026/10/08/what-should-we-tell-our-students/.

Lloyd, S.A. and Sreedhar, S. (2022) ‘Hobbes’s Moral and Political Philosophy’, Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/hobbes-moral/.

Marshall, T.H. (1950) Citizenship and Social Class and Other Essays. Cambridge: Cambridge University Press.

Marx, K. and Engels, F. (1848) Manifesto of the Communist Party. First published February 1848. https://www.marxists.org/archive/marx/works/1848/communist-manifesto/ (accessed 13 October 2026).

Math and AI (2026) A severe misalignment of AI in mathematics. Joint declaration, 11 September. Published 11 September 2026. https://doi.org/10.5281/zenodo.22737750.

Nature Machine Intelligence (2026) ‘Solutions, challenges and rising tensions in AI and mathematics’, Nature Machine Intelligence, 8, p. 857, 23 June. DOI: https://doi.org/10.1038/s42256-026-01269-x.

Norges Bank Investment Management (n.d.) ‘Governance structure’. https://www.nbim.no/en/about-us/about-the-fund/governance-structure/.

OECD (2026) OECD Employment Outlook 2026: Geographic disparities in jobs and incomes. Paris: OECD Publishing. https://www.oecd.org/en/publications/oecd-employment-outlook-2026_7e710f54-en.html.

OpenAI (2026a) ‘An OpenAI model has disproved a central conjecture in discrete geometry’. Published 20 May 2026. https://openai.com/index/model-disproves-discrete-geometry-conjecture/

OpenAI (2026b) Mathematical research repository. GitHub: openai/math. README and history (7 October 2026 entry). https://github.com/openai/math; https://github.com/openai/math/blob/main/history.md (accessed 13 October 2026).

OpenAI (2026c) 'Ten advances in mathematics and theoretical computer science', 1 August. https://openai.com/index/ten-advances-in-mathematics/ (accessed 13 October 2026).

RAISE US (2026) ‘Gina Raimondo and Eric Holcomb launch RAISE US, uniting the nation’s leading employers and bipartisan governors behind American workers’, 25 June. https://www.raiseus.ai/news/release.

Sen, A. (1999) Development as freedom. Oxford: Oxford University Press.

Statistics Netherlands (CBS) (2017). Trends in the Netherlands 2017, section 2.7, ‘Agriculture and horticulture workers’. Statistics Netherlands. https://longreads.cbs.nl/trends17-eng/economy/figures/agriculture/

Sundvall, S. (2024). “Migration and decentralised industrialisation: the development of rural migration in northern Sweden (1850–1950)”, Rural History, 35(1), pp. 71–90. https://doi.org/10.1017/S0956793323000092

Van Meeteren, M. and Smit, M. (2025). “Episodes in Dutch Policy to Develop “Left Behind Places” (1945–2025)”, Tijdschrift voor Economische en Sociale Geografie, 116(5), pp. 544–560. https://doi.org/10.1111/tesg.70048.