Running the Wrong Race: Europe, AI, and the Search for Strategic Leverage

With AI, as in most things, there are rule-makers, and there are rule-takers. This has a tendency to matter more acutely for the rule-takers, as their best-laid plans can easily be waylaid by a change in political leadership, policy direction, or strategic concern elsewhere, and often very far away. My experience working in Singapore is that this is a constant Sword of Damocles hanging over the policy landscape. Still, decisions must be made.

According to the typical discourse within AI policy, there have been three rule-makers. First, the United States, with its vast technology industry, platform dominance, gigantic capital markets, and a laissez-faire model that may move fast and break things, but also builds vast organisations with vaster fortunes. Second, China, which sits in many ways at the opposite end of this spectrum: state-directed innovation, whereby a thousand flowers are left to bloom and the tallest are then heavily invested in, coordinated around, and made central to whole new industries and national capacities. The third is Europe or, more accurately, the European Union, whose strength has long been said to lie in its regulatory capacity, standard-setting, and soft-power projection.

Yet, more and more its bubbling with concerns under the surface. Is the EU on the path to dependence, that would ‘downgrade’ its status as rule-maker? The vulnerability arises from the widening gap between this formal authority and Europe’s material dependence on foreign models, compute, cloud services, platforms, and security infrastructure. A regulator may still determine how a system is deployed while having very little influence over whether that system remains available, how it develops, or which geopolitical conditions govern access to it. Regulation can constrain a provider; it cannot, by itself, create a substitute for that provider. At a certain point, rule-making risks becoming a way of managing dependence rather than overcoming it. Europe, however, would not be a rule-taker in quite the same way as Singapore, nor would its rule-making power suddenly disappear. The EU retains substantial authority to legislate, regulate, and set standards. Its common market gives those rules influence well beyond its borders, and the so-called Brussels Effect remains real (Bradford 2020).

This is the warning dramatised by Europe 2031, an experimental and rather creative policy ‘report’ recently published by a group of European policy researchers and experts. The scenario is right to identify a deficit of material capacity, political commitment, and institutional follow-through. It is less convincing when it allows Silicon Valley’s frontier-model race to stand in for technological destiny, or when ‘Europe’ becomes, in practice, France, Germany, and the European Commission. My argument is therefore less that the scenario is wrong than that it stops too soon. The EU requires not possession of every layer of the AI stack, nor a belated attempt to become California with better trains, but sufficient leverage to choose, negotiate, refuse, and recover. The central question is not simply whether Europe can run faster, but whether it is running the right race.

Europe 2031 is less a conventional policy report than a novella-manifesto. Its purpose is not only to analyse The EU’s position in AI, but to paint a less abstract picture of where the continent may be heading. As a potential future, it offers a dire warning, seemingly in an attempt to make that future feel real enough that action in the present becomes urgent.

The scenario follows Caroline Dubois, a European Commission official who gradually realises that Europe is misreading both the speed and strategic significance of AI. Beginning with an initial, and misplaced, celebration of DeepSeek in 2025 as evidence that the EU might cheaply catch up with the United States and China, it traces a widening gap in compute, capital, frontier-model capability, and institutional adoption. American AI systems become increasingly essential to cybersecurity, productivity, and scientific research, while Washington progressively restricts European access. China, meanwhile, consolidates its lead in robotics and industrial AI. European leaders consistently fail to read the room.

The EU eventually embraces ‘AI sovereignty’, but predominantly through underfunded domestic champions, procurement mandates, and regulatory pressure rather than by building infrastructure or bargaining power. The resulting productivity gap erodes Europe’s tax base, welfare systems, and political cohesion. By 2031, the EU’s only substantial strategic asset—ASML—is caught between American and Chinese coercion, leaving its leaders to choose between two forms of dependency rather than exercise meaningful sovereignty.

The scenario’s central analytical claim is that this failure would arise less from regulation itself than from a mismatch between the pace of technological change and Europe’s fragmented, risk-averse political machinery. It distinguishes symbolic ownership—having an EU model or an EU cloud—from material leverage: compute under European jurisdiction, control over supply-chain bottlenecks, credible alliances, and the capacity to adopt AI throughout the economy. It also argues that political caution can become self-defeating: that preserving existing labour protections, permitting arrangements, and procurement practices may ultimately destroy the social model those rules were intended to protect. The epilogue complicates this rather technocratic prescription by acknowledging that drastic reform would impose real democratic costs and might trigger its own backlash.

The final lesson is therefore partly political and partly narrative. Europe lacked not only infrastructure and speed, the scenario suggests, but a persuasive account of what an AI-enabled European future was for. This matters, because Europe 2031 is not simply carrying an argument in an unusually entertaining container. The speculative form is part of the argument.

AI policy has long been shaped through speculative futures, often presented through some form of speculative fiction. Vernor Vinge’s technological singularity, Nick Bostrom’s superintelligence, and the more recent book by Eliezer Yudkowsky and Nate Soares are rather different interventions, but all use imagined futures to act upon the present (Vinge 1993; Bostrom 2014; Yudkowsky and Soares 2025). They tell us what to fear, what to value, which actors ought to be trusted, and what sacrifices suddenly appear reasonable. Europe 2031 belongs to this broader family, even if its protagonists are Commission officials rather than recursively self-improving machines.

These futures are not neutral portrayals. They operate as sociotechnical imaginaries: collectively held and institutionally stabilised visions of social life made possible through science and technology (Sismondo 2020; Suchman 2022). Such imaginaries—and the stories through which they become tangible—frame what counts as urgent, relevant, or realistic, and which trade-offs begin to seem necessary. The authors of Europe 2031 make no secret of their intention to influence the EU’s AI agenda. The scenario is therefore interesting not only because of what it says—or, indeed, does not say—but because of how it teaches the reader to see the problem.

This is why treating the document as a prediction, and then arguing over whether every event is likely, rather misses the point. My own rule of thumb is that once one moves beyond second-order ‘ifs’, one is firmly in speculative territory. The assumptions about politics, society, and technological change then matter more than whether the imagined Commission meeting occurs in March or September. A scenario connects a chain of contingent developments, makes one version of the future vivid, and relies on the reader to experience that vividness as plausibility. Its power comes precisely from turning uncertainty into a story with momentum.

Scenarios, in other words, do not merely describe possibilities. They organise them. The authors may not control every interpretation of their story, but they retain considerable control over where the story begins, which actors receive agency, what counts as a crisis, and which exits remain visible. If the future is framed as an AGI- and frontier-laboratory race, then scale, compute, deregulation, and centralised mobilisation quickly become the obvious answers. If the EU’s AI future is instead framed around strategic resilience, industrial pluralism, and democratic capacity—the kinds of things routinely lauded as ‘European values’ (Bercusson 2009; Douglas-Scott 2011)—the available answers begin to look rather different.

Prophecy, in this sense, tells us at least as much about how the present is being seen as it does about what the future will contain (Ardener 2017 [1989]). Europe 2031 offers a compelling diagnosis of Europe’s present institutional weakness. In making one particular technological future feel inevitable, however, it also risks reproducing some of the limits Europe needs to overcome. A review of the scenario must therefore ask two questions at once: is the warning useful, and what political work is being done by the future through which that warning is delivered?

My key takeaway from the scenario is its diagnosis of the EU. The problem the EU faces is not merely one of regulatory caution, as has often been asserted (Bradford 2024; Castro 2023; Leparmentier 2024; Madhani and Adamson 2025). It is a problem of political will, institutional fragmentation, and underpowered industrial strategy. Much of what the EU tends to lean on—big announcements with lofty ideals, earth-shattering investments and visions that will transform the region’s entire industrial landscape, always about five years from now—confuses announced money with real mobilisation. AI sovereignty, whatever that means in practice, is much easier to declare than to build. It is also an ongoing political and institutional project, rather than something one constructs and thereafter simply continues to ‘have’.

Speaking to people working in Europe and alongside the EU, together with my own admittedly more limited experience, reinforces this diagnosis. Many of the relevant stakeholders already recognise the central bottlenecks: data, trust, energy, compute, scale, talent, cybersecurity, procurement, and adoption. The unresolved questions are procedural and organisational. Who is responsible for committing the necessary resources? Who controls them? Who can be held accountable for delivery when the grand announcement encounters a power grid, a procurement rule, a national ministry, or an election?

Although the EU has emphasised rules—and, as someone working in AI governance, I welcome these initiatives—rules without capacity quickly becomes dependence management. The ability to make a supplier report risks is not the same as the ability to substitute for it, maintain access during a political dispute, or continue operating if the supplier withdraws. Enforcing rules can improve behaviour and accountability; it does not necessarily create compute, cloud capacity, technical expertise, bargaining power, or an alternative provider. The EU’s regulatory power is real, but the authority to govern a technology and the capacity to shape or sustain that technology are not interchangeable.

This is where the report’s language of rule-making and rule-taking needs to be sharpened rather than discarded. Europe is not simply moving from one category to the other. It remains a rule-maker in legal and regulatory terms while becoming a technology- and infrastructure-taker in strategically important parts of the AI stack. The danger is that material dependence progressively hollows out formal authority. Rules can still bite, but what they bite into—and whether Europe can live without it—may be determined elsewhere.

The scenario is also right to insist that implementation is not a technical stage that naturally follows political agreement. Infrastructure must be permitted, financed, connected to energy, staffed, procured, and used. Domestic firms need customers as well as grants. Public agencies need technical personnel and organisational authority, not merely pilots or ethics principles. Serious mobilisation also creates winners, losers, and opportunity costs: money directed to compute cannot simultaneously fund another priority; faster permitting imposes local costs; procurement designed to support European providers may be more expensive in the short term; labour-market adaptation raises the decidedly political question of who absorbs disruption. Commitment begins when political actors specify not only the desired destination, but who will provide the resources, accept the risks, and remain accountable when delivery becomes difficult.

My disagreements with Europe 2031 are not intended to disprove its overall diagnosis. They concern the assumptions built into the future it constructs—and, consequently, the range of responses the scenario makes politically imaginable. These can be condensed into three challenges.

While Europe 2031 correctly warns against European complacency, it often does so by accepting Silicon Valley’s own account of AI progress: fast, frontier-led, compute-driven, AGI-adjacent, and historically inevitable. The scenario largely treats progress as though it moves along one main line, running from frontier laboratories and their demand for ever-greater quantities of compute towards increasingly general systems, with something resembling AGI waiting further along the trajectory.

It would be ridiculous to deny the advances of recent years, or the importance of frontier models and compute. Nevertheless, the proposed progression from contemporary AI to artificial general intelligence and then artificial superintelligence remains a live debate, not an inevitable technological pipeline. This matters because the scenario sometimes risks equating reasonable scepticism with denialism. Hallucinations, cybersecurity, data leakage, organisational friction, market sustainability, environmental cost, and the difficulty of adoption are not merely ‘EU cope’. Capability is not the same as reliable capability, and neither necessarily translates into useful institutional adoption—nor, in every conceivable setting, should it.

More fundamentally, AI is not one technology, one industry, or one race. It is a family of models, infrastructures, applications, robotics systems, industrial practices, scientific tools, public services, and institutional arrangements. The way the race is defined determines where agency can be found. If the only meaningful contest is the construction of increasingly general frontier models, Europe is simply behind. If AI comprises multiple contests, Europe may remain behind in some while retaining considerably more choice over where and how it competes.

The race metaphor does political work here. Once frontier-model capability becomes the principal measure of AI success, compute scale, laboratory concentration, and speed become the natural indicators of seriousness. Public-sector reliability, industrial productivity, scientific usefulness, energy efficiency, or democratic control are relegated to the apparently secondary matter of deployment, after the ‘real’ technological contest has been decided. The metaphor does not merely describe Europe’s position; it determines which positions count.

The EU may not be well positioned to win a head-on frontier-model race against American hyperscalers on their preferred terms. It may be considerably better positioned in industrial and physical AI, robotics, scientific AI, energy systems, healthcare, semiconductors and photonics, AI assurance, trusted data infrastructure, and specialised systems built around high-quality institutional data. These are not consolation prizes. They are domains in which reliability, sectoral knowledge, established institutions, and legitimacy matter, and in which Europe already possesses industrial or institutional strengths. European policymakers should not ignore Silicon Valley, but nor should they outsource their technological imaginations to it.

Perhaps this stood out to me particularly because I come from a much smaller EU country, but the scenario is ostensibly about the European Union while almost all meaningful action takes place through France, Germany, and the European Commission. Politically, this is understandable: these are powerful actors whose decision-makers the scenario is intended to reach. Strategically, however, it is limiting.

France and Germany clearly matter, and any coherent strategy would benefit enormously from having both on board. Yet the EU’s strengths do not reside solely—or perhaps even primarily—in its largest member states. Smaller states have long had to navigate limited scale, energy constraints, technological dependence, public-sector digitalisation, organisational transformation, and geopolitical exposure without being able to dictate the terms. Countless examples have shown that being a rule-taker does not mean having no agency. It does mean navigating stormier seas, identifying specific sources of leverage, and making difficult trade-offs in pursuit of pragmatic gains. These are not peripheral skills in the emerging AI landscape—but remain unrecognised by these ‘bigger powers’ (Luttwak 2012).

Estonia matters for digital government; the Nordic countries for trust, public services, energy, and innovation capacity; the Netherlands for semiconductors and strategic infrastructure; Ireland for platform presence and regulatory leverage. Poland is curiously absent despite becoming increasingly central to the EU’s geopolitical future. The common market is itself a considerable strategic asset, but only if it is understood as more than a Franco-German extension. Europe’s diversity can certainly produce fragmentation, but it can also provide a wider repertoire of practical experiments. Its problem is not simply that it contains too many systems, it is that it has repeatedly failed to translate those differences into coordinated strength.

A more credible strategy would therefore be deliberately diffused. Different countries and regions should be able to lead where they possess credibility, capacity, and urgency, with coalitions forming around particular capabilities rather than waiting for all twenty-seven member states to move in perfect synchrony on every issue. The European Space Agency offers a partial precedent: alongside mandatory activities, its optional programmes allow states to invest according to their interests, while geographical return directs contributions towards relevant domestic industries, research institutes, and universities. The analogy should not be stretched too far—AI presents rather different commercial and institutional problems—but it demonstrates that European cooperation need not mean either complete centralisation or twenty-seven disconnected national strategies.

Finally, Europe 2031 says it wants to preserve the EU’s social model, but several of its implied solutions suggest that Europeans must become more American in order to survive American dominance. Deregulation, weaker labour protections, looser copyright rules, faster permitting, and a more hyperscaler-friendly industrial policy appear as the price of remaining relevant. Some European sacred cows may indeed deserve a nervous glance towards the abattoir. These choices should not, however, be presented as a simple fork between Americanisation, Sinification, and obsolescence.

The United States and China are each playing to their own strengths. The United States draws on deep capital markets, hyperscalers, platform dominance, venture risk, and military-industrial integration. China combines manufacturing capacity, state-backed coordination, physical infrastructure, and industrial deployment. The scenario recognises that Europe is attempting something closer to the Chinese approach without the institutional capacity or political coordination required to make it work, yet its implied alternative remains largely American. This smuggles a political-economic preference into what otherwise appears to be a technological necessity. The relevant question is not simply which regulations should be retained or relaxed, but which social and political arrangements Europeans are attempting to sustain through AI strategy. A serious European response cannot simply be ‘Silicon Valley, but with GDPR paperwork’. Nor can it abandon the European social model in the name of saving it without confronting the contradiction involved.

This is also where sovereignty needs more careful treatment. AI sovereignty should not mean owning everything. That is neither realistic nor necessarily desirable. Sovereignty is not autarky; it is the possession of enough capacity, bargaining power, and resilience to avoid dependency traps. It is the ability to choose between partners, negotiate terms, protect sensitive data, maintain critical services, refuse unacceptable conditions, and recover when relationships fail. If Europe 2031 succeeds in making dependence visible, the next step is to imagine sovereignty from the position the EU actually occupies.

The recommendations in Europe 2031 are not presented as a discrete policy list, but baked into the scenario and its epilogue. In broad terms, it calls for emergency mobilisation around compute, energy, infrastructure, talent, and adoption; faster permitting; partnerships that anchor hyperscaler infrastructure on European soil; investment in robotics and industrial AI; labour-market adaptation; coalitions among middle powers; and a more positive account of why the resulting disruption would be worthwhile.

These recommendations are useful because they refuse the comforting fiction that Europe can regulate its way out of strategic dependence. Yet urgency is not the same thing as strategy. If the EU responds only by attempting to become a late, smaller, and more constrained version of Silicon Valley, it risks losing twice: first by failing to catch up on American terms, and second by weakening the political and social model it claims to defend.

The first task is to convert urgency into commitment—here the report and I are in full agreement. Europe needs institutions, funding structures, procurement pathways, and political agreements that bind actors to delivery over time: real funding rather than creative accounting; clear ownership of strategic priorities; shorter infrastructure and procurement timelines; and public investment tied to deployment rather than another cycle of pilots. Energy, compute, talent, cybersecurity, industrial policy, and public-sector adoption cannot remain separate files. Data centres depend on grid capacity; firms cannot scale without customers and infrastructure; public institutions cannot govern systems they lack the technical capacity to understand.

The point is not speed for its own sake. Speed without strategy becomes panic, while strategy without commitment becomes theatre. Europe needs both the capacity to decide and the machinery to follow through. That also means being honest about opportunity costs, rather than announcing every priority simultaneously and hoping that the conflicts will disappear somewhere between Brussels and implementation.

Second, the EU should stop treating AI as though it were one race with one finishing line. It does not need ‘its own OpenAI’ as a prestige project so much as capabilities that others need, trust, and cannot easily replace. Physical and industrial AI, robotics, scientific research, energy-efficient computation, healthcare, semiconductors and photonics, assurance and evaluation, public-sector systems, and trusted data-sharing infrastructure provide a plausible starting portfolio.

The precise portfolio requires evidence and choices that no slogan can substitute for. Which systems are critical to public continuity or security? Where does Europe possess data, engineering expertise, manufacturing capacity, or trusted institutions that competitors cannot rapidly reproduce? Which capabilities would create bargaining power elsewhere in the stack? Which investments remain defensible on public-value grounds even if they never produce a globally dominant firm? A credible strategy must name what Europe will not attempt to build as readily as it announces what it will.

Third, Europe should not respond to dependence on the United States through fantasy autarky. Many technologically capable countries face a similar problem: they cannot match the scale of the United States or China, but nor can they afford to become passive consumers of systems built elsewhere. The United Kingdom, Japan, South Korea, Canada, Norway, Singapore, and Australia, where politically feasible each control or influence parts of the AI stack. Cooperation on compute access, supply-chain resilience, cybersecurity, evaluation, industrial AI, standards, trusted data, and procurement could turn fragmented assets into shared leverage.

Such partnerships require a more mature approach to both the United States and China. Europe should not be naïve about Chinese state power, surveillance, industrial strategy, security risks, or coercive capacity. Nor should it treat dependence on the United States as inherently safe because ‘better the Devil you know’. American control over central financial and technological infrastructures has repeatedly allowed Washington to convert interdependence into geopolitical leverage (Farrell and Newman 2019; 2023). In cloud infrastructure, semiconductors, communications, defence, and advanced AI, dependence on the United States remains dependence.

This does not make every dependency equivalent or turn China into an uncomplicated partner. China will be a competitor in some areas and a substantial security risk in others. It may nevertheless offer useful lessons in industrial AI, physical deployment, infrastructure coordination, open models, and state-backed scaling. Europe need not copy China any more than it should copy the United States. It should be capable of learning from both without mistaking either for its own model.

Finally, Europe needs more than warnings. Europe 2031 is powerful because it makes decline, dependence, irrelevance, and political fragmentation vivid. Fear can generate attention, but it cannot sustain a political project. Europeans need a positive account of what AI capacity is for: better public services, resilient industries, stronger democratic institutions, scientific discovery, cleaner energy systems, safer infrastructure, improved healthcare, and labour-market adaptation that does not treat workers as collateral damage.

This returns us to the politics of speculative futures. A scenario that imagines only panic and belated imitation makes panic and belated imitation appear sensible. A scenario organised around leverage, plural technological pathways, and a recognisably European political project would make a different set of interventions imaginable. Europe does not only need more AI capacity; it needs a future worth building that capacity for. The real danger is not that Europe fails to win the American AI race, but that it never asks whether this was the right race to enter.

The debate matters beyond Europe because the underlying problem is shared by many states and regions. Very few can match American capital depth or Chinese industrial coordination. The danger is that everyone else becomes drawn into someone else’s race: frontier models, data-centre scale, chip access, platform dependence, and geopolitical alignment presented as technological inevitabilities.

This is particularly relevant to Southeast Asia. ASEAN is not the EU and does not seek to become it. It lacks equivalent supranational authority, while its members differ even more sharply in size, political organisation, economic structure, and strategic ambition. Its challenge is nevertheless related: how to avoid becoming a passive deployment zone for systems, infrastructures, standards, and assumptions developed elsewhere.

ASEAN’s leverage will not resemble European regulatory power or American scale. It may instead emerge from logistics, finance, manufacturing, digital infrastructure, public administration, multilingual societies, regional trade networks, and considerable experience navigating between larger powers. These are not peripheral to AI strategy. They are precisely the areas in which more grounded and defensible forms of capability might emerge, whether through public-sector systems, logistics and trade facilitation, model evaluation, cybersecurity cooperation, or sector-specific applications in healthcare, education, agriculture, finance, and urban governance.

Strategic weakness rarely arrives as a single decision. It hardens through repeated choices of cloud provider, procurement contracts that cannot be exited, imported standards, absent local expertise, and public institutions that understand critical systems less well than the vendors selling them. By the time dependence becomes politically visible, many alternatives may already have disappeared. Europe’s predicament is therefore useful to Southeast Asia not as a model to copy, but as an early warning about the cumulative politics of apparently technical choices.

Europe 2031 succeeds as a warning. It makes European decline feel concrete and forces a conversation about AI, infrastructure, political will, and strategic dependence that Europe should already have been having. Its central diagnosis is difficult to dismiss: the EU risks mistaking rules for capacity, announcements for commitment, and symbolic sovereignty for actual leverage. Europe is not simply a rule-taker, because it can still write rules with global consequences. Yet reliance on foreign models, compute, cloud services, platforms, and security infrastructure risks leaving those rules suspended over a technological system whose development and conditions of access are determined elsewhere.

The appropriate response is neither regulatory retreat nor fantasy autarky. Europe must convert urgency into commitment, build around differentiated strengths, and construct partnerships that preserve room to choose, negotiate, refuse, and recover. It must also resist the temptation to treat one speculative future as technological destiny. AI is not one technology, one race, or one future. Europe’s task is not simply to move faster, but to decide what kind of technological future it is trying to build—and then acquire enough material leverage to make that decision matter.

Ardener, Edwin. 2017 [1989]. “The Voice of Prophecy—Further Problems in the Analysis of Events”. The Voice of Prophecy: And Other Essays (Malcolm Chapman, Maryon McDonald, & Kirsten Hastrup, eds.). Berghan Books. 134-154.

Bercusson, Brian. 2009. “The Lisbon Treaty and Social Europe.” ERA Forum 10: 87–105.

Bostrom, Nick. 2014. Superintelligence: Paths, Dangers, Strategies. Oxford University Press.

Bradford, Anu. 2020. The Brussels Effect: How the European Union Rules the World. Oxford University Press.

Bradford, Anu. 2024. “The False Choice Between Digital Regulation and Innovation.” Northwestern University Law Review 119: 377.

Castro, Daniel. 2023. “The EU’s AI Act Is Premature.” ITIF.

Douglas-Scott, Sionaidh. 2011. “The European Union and Human Rights after the Treaty of Lisbon.” Human Rights Law Review 11(4): 645–682.

Farrell, Henry, and Abraham L. Newman. 2019. “Weaponized Interdependence: How Global Economic Networks Shape State Coercion.” International Security 44(1): 42–79.

Farrell, Henry, and Abraham L. Newman. 2023. Underground Empire: How America Weaponized the World Economy. Macmillan.

Leparmentier, Arnaud. 2024. “Reid Hoffman, ‘Mystical Atheist’ and Artificial Intelligence Guru.” Le Monde.

Luttwak, Edward N. 2012. The Rise of China vs. the Logic of Strategy. Harvard University Press.

Madhani, Aamer, and Thomas Adamson. 2025. “JD Vance Rails Against ‘Excessive’ AI Regulation in a Rebuke to Europe at the Paris AI Summit.” Associated Press.

Sismondo, Sergio. 2020. “Sociotechnical Imaginaries: An Accidental Themed Issue.” Social Studies of Science 50(4): 505–507.

Suchman, Lucy. 2022. “Imaginaries of Omniscience: Automating Intelligence in the U.S. Department of Defense.” Social Studies of Science 53(4): 761–786.

Vinge, Vernor. 1993. “The Coming Technological Singularity: How to Survive in the Post-Human Era.” Vision-21 Symposium.

Yudkowsky, Eliezer, and Nate Soares. 2025. If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All. Little, Brown and Company.