The AI race works like the trap Scott Alexander called Moloch: the leading developers say a slower, coordinated pace would be safer, yet none has committed to slowing down alone while rivals keep going, and they have now written that reasoning into their own safety policies. In February 2026 Anthropic rebuilt its Responsible Scaling Policy around what it calls a collective action problem. OpenAI's Preparedness Framework lets the company adjust its required safeguards if a rival releases a high-risk system without comparable ones, and Google DeepMind's framework says some of its mitigations work for society "only if all relevant organizations provide similar levels of protection".
Moloch is a trap that needs no villain
Alexander's essay "Meditations on Moloch", published on Slate Star Codex on 30 July 2014, used Moloch as a name for what he called multipolar traps. Once one competitor gains an edge by sacrificing a value everyone shares, the rest must sacrifice it too or be "outcompeted and replaced by the less scrupulous", and all end up worse off although each followed its own incentives; arms races are one of his examples. Alexander argued that the only lasting way out was a Gardener, an intelligence that optimises for human values, which superintelligence might put within reach, and he warned that if several competing entities built one at once, "we would be super-doomed".
Researchers at Oxford's Future of Humanity Institute had already formalised the AI case. In a deliberately simplified model, first circulated in 2013 and published in AI & Society in 2016, Stuart Armstrong, Nick Bostrom and Carl Shulman showed teams racing to build a powerful AI skimping on safety to finish first. More teams, and more enmity between them, could raise the risk of disaster; less intuitively, so could better information about each other's capabilities.
The labs have written the race into their safety policies
Anthropic published its Responsible Scaling Policy in September 2023, hoping to start what it calls a "race to the top", and OpenAI and Google DeepMind followed with comparable frameworks within eight months. The current versions show the race shaping all three. Section 4.3 of OpenAI's Preparedness Framework, as revised on 15 April 2025, says that if another developer releases a system with high or critical capability without comparable safeguards, OpenAI could adjust its own, provided it says so publicly, judges that overall risk does not meaningfully rise, and stays more protective than the rival "in order to avoid a race to the bottom on safety". We found no record of its use.
Google DeepMind's Frontier Safety Framework said in February 2025 that the company's adoption of its own protocols "may depend on" whether others across the field adopted similar ones. Versions 3.0 (September 2025) and 3.1 (April 2026) dropped that sentence but still say certain mitigations protect society only if every relevant developer matches them, and let a model's recommended security level be relaxed if comparably capable public models have weaker security.
Anthropic went furthest. The third version of its policy, effective 24 February 2026, says earlier versions committed the company to acceptable risk levels regardless of what other developers did, and argues that if one developer paused while others pressed on without strong mitigations, "the developers with the weakest protections would set the pace". An appendix, "Commitments Related to Competitors", promises delays in two cases: if Anthropic has a clear lead, until it can make a strong case that catastrophic risk is contained or the lead disappears; and if every competitor at the frontier can already make that case, until Anthropic matches them. Jared Kaplan, its chief science officer, told TIME that unilateral commitments made little sense "if competitors are blazing ahead", while denying the change was a U-turn. The policy concedes that the scenario motivating it "has not yet arisen", and an April 2026 revision says Anthropic remains free to pause in circumstances the appendix does not cover.
Each clause is defensible alone, since a safeguard kept by one developer may buy the world little. Together they tie every lab's safety floor to the others', the coupling the racing model predicts.
The people running the labs describe the same bind
In May 2023 the chief executives of OpenAI, Google DeepMind and Anthropic, Sam Altman, Demis Hassabis and Dario Amodei, signed the Center for AI Safety's one-sentence statement that mitigating the risk of extinction from AI should be "a global priority". OpenAI's Charter of April 2018, still on its website, warned of "a competitive race without time for adequate safety precautions" and promised to stop competing with, and start assisting, a safety-conscious project that got close to AGI first; the principles Altman published in April 2026 do not repeat that promise, saying instead that OpenAI expects periods when it must work with governments and "other AGI efforts" on safety before proceeding further.
At Davos in January 2026, Amodei said he wished for the five to ten years Hassabis expects before AGI-level systems, rather than the one to two he thinks likely, a spread typical of AGI forecasts. When the moderator, The Economist's editor Zanny Minton Beddoes, replied that he could just slow down, he said rival states were building the same technology at a similar pace and "it's very hard to have an enforceable agreement", according to the Forum's transcript. Hassabis said a slightly slower pace might be good but would require coordination. In a July 2026 essay he wrote that "we are locked in an extremely intense, multilayered commercial and geopolitical race" and proposed a US standards body that could, if necessary, coordinate a slowdown among frontier labs.
"Pacing the Frontier", published on 28 July 2026 and signed, according to its website, by 1,386 employees of frontier AI companies, among them Amodei, Kaplan and OpenAI's chief scientist Jakub Pachocki, says each company and country is "under intense competitive pressure not to unilaterally slow" an acceleration that automated AI research could bring. It asks Washington to back an international effort to build tools for pacing the frontier: the means to slow down rather than a slowdown, as our analysis of the statement notes.
Amodei went further in September. In an essay published on 12 September 2026, he wrote that "we must slow the pace at which we improve the capabilities of AI models". The step Anthropic took on alone was to give outside evaluators employee-like access; limits on the pace itself, he wrote, would need coordination among companies, which for antitrust reasons would need the government's support. Altman replied that "we need to pace the frontier" and that OpenAI would match the evaluator pledge, and Musk posted "Dario is right". The next day President Trump dismissed the calls, saying "whoever wins AI, wins". On 18 September four paying subscribers sued Anthropic, OpenAI, SpaceXAI and Google, alleging that the companies had agreed to slow down in breach of antitrust law.
The money raises the price of stopping
The sums committed since 2025 make a unilateral pause more expensive every quarter. OpenAI, SoftBank, Oracle and MGX announced Stargate in January 2025, a new company meant to invest $500 billion in AI infrastructure for OpenAI over four years and "secure American leadership in AI". Alphabet's second-quarter 2026 results show $44.9 billion spent on property and equipment in three months, double the figure a year earlier, and negative free cash flow. Meta expects capital expenditure of $130 billion to $145 billion in 2026. Anthropic raised $30 billion at a $380 billion valuation in February 2026 and then $65 billion at $965 billion in May; its finance chief said the money would help it "stay at the research frontier".
There are signs the pressure reaches testing. The Financial Times reported in April 2025, citing eight people familiar with the process, that OpenAI was giving staff and outside testers days rather than months to assess new models. OpenAI told the paper the shorter timeline reflected efficiencies and automation.
Coordination has produced disclosure, not brakes
The Future of Life Institute's open letter of 22 March 2023, signed by Elon Musk, Yoshua Bengio and Stuart Russell among more than 30,000 others, asked all labs to pause training systems more powerful than GPT-4 for at least six months; no leading lab announced a pause, as our piece on the pause debate recounts. In July 2023 seven companies made voluntary commitments to the White House, including security testing before release. That November, 28 countries and the EU, the United States and China among them, agreed the Bletchley Declaration, which warned of "serious, even catastrophic, harm". At Seoul in May 2024, 16 companies, including China's Zhipu AI and xAI, signed commitments to publish safety frameworks and, in the extreme, "not to develop or deploy a model or system at all" if risks could not be kept below their own thresholds.
What these efforts produced was disclosure. METR, a research nonprofit, now lists published frontier safety frameworks from 12 companies, and California's SB 53, signed in September 2025, requires large frontier developers to publish one. It leaves the thresholds, and the decision to stop, with the companies. The published frameworks are also the documents that now carry competitor clauses.
Governments have joined the race
At the Paris AI Action Summit in February 2025 the United States and the United Kingdom did not sign the closing statement, which China signed, and Vice President JD Vance said, as TechCrunch reported, that the AI future "is not going to be won by hand-wringing about safety". The White House's AI Action Plan of July 2025, titled "Winning the Race", opens by declaring that the United States "is in a race to achieve global dominance". A December 2025 executive order setting up a task force to challenge state AI laws says the country is "in a race with adversaries for supremacy". China proposed a World AI Cooperation Organization in July 2025, according to its foreign ministry.
Amodei uses the same frame. After the Chinese lab DeepSeek released its R1 reasoning model in January 2025, he argued that export controls would decide whether the world became "unipolar or bipolar" in AI, and at Davos he said that without chip sales to China the race would become "competition between me and Demis", which he was confident they could work out. The administration went the other way: in January 2026 the Commerce Department began reviewing licences to export Nvidia's H200 and similar chips to China case by case, relaxing the chip export controls Amodei had urged it to tighten.
The trap may be less closed than it looks
There are credible arguments that the Moloch framing overstates the bind. Katja Grace, of the research group AI Impacts, argued in a December 2022 essay that it is "very non-obvious that we are in or near an arms race". If winning may mean a system nobody controls, she argued, racing need not be anyone's best strategy, and she pointed to past restraint such as the Asilomar conference on recombinant DNA.
The clearest sign that an outside brake can work came in June 2026. Ten days after an executive order called for a voluntary scheme giving the government early access to frontier models with advanced cyber capabilities, the US government applied export controls to Anthropic's newly released Claude Fable 5 and Claude Mythos 5. Anthropic suspended both models for all users; the controls were lifted on 30 June, the company said. On 26 June OpenAI limited a preview of its GPT-5.6 models at the administration's request, while writing, in a post quoted by TechCrunch, that such a process should not become "the long-term default". Both labs complied. The benchmark that decides which models the June order covers is classified, and the order, as our piece on the voluntary framework noted, creates no licensing requirement.
Alexander counted coordination among the forces that hold multipolar traps in check. Since 2023 the labs have converged faster than governments: they built similar frameworks within a year, while the US government, which endorsed the Bletchley Declaration, now calls AI a race to be won. Because the labs' policies are written to track the rest of the field, a constraint that holds will have to come from outside them. June showed that one can be applied within days; whether it can be applied under written rules, and across borders, is the open question.
Correction, 14 October 2026: An earlier version of this article said that a leading lab announced a major model release the week after the Future of Life Institute's pause letter was published. We found no such release; OpenAI had released GPT-4 eight days before the letter. The earlier version also attributed to Sam Altman a public statement that his company might be building one of the most dangerous things in human history, for which we could find no source, and it named Baidu and Zhipu AI as China's main AI developers without a source, leaving out DeepSeek, whose R1 model had drawn worldwide attention in January 2025. Those passages have been corrected or removed.