The pause demanded in a 2023 open letter never happened, but the argument behind it has spread from outside the AI labs to inside them: in September 2026 the chief executives of Anthropic and OpenAI said frontier development should slow down. The Future of Life Institute's letter, dated 22 March 2023, asked all labs to "immediately pause for at least 6 months" the training of systems more powerful than GPT-4; by October 2026 FLI's website listed more than 30,000 signatures. In August 2026 OpenAI disclosed that it had paused some frontier training for two weeks. On 12 September Anthropic's Dario Amodei called for a coordinated slowing of capability gains, and OpenAI's Sam Altman agreed within hours. Two days later President Donald Trump called fears of AI destroying humanity a hoax, and the White House has since promoted industry self-regulation.
The 2023 letter asked for six months and got none
The letter wanted a pause that was public, verifiable and joined by all key actors, and said governments should step in with a moratorium if one could not be enacted quickly. The time was to go on shared safety protocols audited by independent outside experts; the letter stressed that it was not seeking a pause on AI development in general. It quoted OpenAI's own remark, a month earlier, that it might one day be important to limit the growth of training compute, and replied: "That point is now." Signatories included Yoshua Bengio, Stuart Russell, Elon Musk, Steve Wozniak and Yuval Noah Harari. A PDF of the letter that FLI created on 5 May 2023 listed 27,565 signatures.
Eliezer Yudkowsky declined to sign, writing in Time on 29 March 2023 that the letter understated the danger and asked for too little; he wanted an indefinite, worldwide moratorium on large training runs. Altman, whose company had just released GPT-4, told an audience at MIT in April that the letter was "missing most technical nuance about where we need the pause", The Verge reported. On 30 May 2023 he, Demis Hassabis of Google DeepMind, Geoffrey Hinton and Bengio did sign a one-sentence statement from the Center for AI Safety, which said mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war. It asked for no pause. A year after its letter, the Future of Life Institute conceded that a pause had not materialised, saying developers had shown that their "primary commitment is to speed".
Four different proposals share one name
The mildest is the conditional pause: stop if a model crosses a danger threshold before safeguards are ready. Anthropic's first Responsible Scaling Policy, in September 2023, said it implicitly required the company to pause training more powerful models temporarily if scaling outstripped its safety procedures. At the AI Seoul Summit on 21 May 2024, sixteen companies, among them OpenAI, Google, Meta, xAI and China's Zhipu.ai, committed, in the extreme, "not to develop or deploy a model or system at all" if they could not keep risks below their own thresholds. The 2023 letter was the second kind, a fixed-term moratorium above a capability line.
The third is prohibition until conditions are met. PauseAI, founded in Utrecht in May 2023 by Joep Meindertsma, wants a global pause on training the most powerful general AI systems until they can be built safely and kept under democratic control, secured by a treaty. The Future of Life Institute's Statement on Superintelligence, released on 22 October 2025 with more than 800 signatures, calls for a "prohibition on the development of superintelligence" until there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in. Early signers included Hinton, Bengio, Wozniak, Steve Bannon, Susan Rice and several China-based AI researchers, NBC News reported. By mid-October 2026 the statement's site counted about 145,000 signatures, roughly half of them gathered through a parallel petition on Ekō.
The newest is pacing: slowing progress by agreement without stopping it. In July 2026 more than 1,200 employees of frontier AI companies signed a statement asking Washington to back an international effort to build tools for pacing frontier AI; by mid-October its website listed 1,386 signatories. Amodei's September essay insisted that pacing "does not mean halting model training or technical progress".
The objections come from named critics, and from the labs
The first is practical. Andrew Ng, founder of DeepLearning.AI, called the 2023 proposal "a terrible idea", arguing that better models would help people through new applications in education, healthcare and food. There was "no realistic way" to stop every team scaling up language models unless governments stepped in, he wrote, and it would be anti-competitive for governments to pause technologies they do not understand. The AI researcher Nora Belrose added a technical objection in September 2023: most pause proposals, including the letter, left hardware development untouched, so lifting a pause all at once would release a hardware overhang, a sudden jump in capability that could bring the kind of fast take-off she judged most dangerous. The same reasoning about compute underlies chip export controls as a brake.
The second is competition. Mark Zuckerberg wrote on 10 August 2026 that any policy slowing American model releases, "even by a month", could put American leadership at risk while foreign models raced ahead. Trump said a conspiracy against AI and data centres was pleasing only China, and Beijing dismissed the American warnings as fearmongering, NBC News reported. Amodei grants part of the point: his essay says slowing within democracies is limited by their lead over China, and that a full global pause, though worth floating, is unlikely soon because defection would be tempting and verification hard.
The third says the framing points at the wrong problem. Timnit Gebru, Emily M. Bender and Angelina McMillan-Major, co-authors of the "Stochastic Parrots" paper, wrote that the 2023 letter's sensible proposals were "overshadowed by fearmongering and AI hype" and ignored present harms such as worker exploitation, data theft and concentrated power. Arvind Narayanan and Sayash Kapoor, in their April 2025 essay "AI as Normal Technology", call superintelligence "incoherent as usually conceptualized" and argue that drastic interventions premised on the difficulty of controlling it would make things much worse if AI proves to be a normal technology; they favour building resilience.
The fourth is written into the labs' own rules. When Anthropic rewrote its scaling policy in February 2026, it dropped its commitment to keep its own risks acceptable regardless of what competitors did, explaining that if one developer paused while others pressed on, "the developers with the weakest protections would set the pace". The current version commits to delaying development only if Anthropic has a significant lead or every competitor has strong safety measures, though it adds that the company would "strongly consider" pausing in other cases. OpenAI's April 2025 revision of its safety framework said it "may adjust" its requirements if a rival released a high-risk system without comparable safeguards, after first confirming that the risk had really changed. This is the race dynamic set down in policy.
The fifth is legal. On 18 September 2026, four people who pay for Claude, ChatGPT, Grok or Gemini filed a proposed class action against Anthropic, OpenAI, SpaceXAI and Google in the federal court for the Northern District of California, alleging that the companies had illegally agreed to slow AI development, the Associated Press reported. The complaint centres on 12 September, when Elon Musk, SpaceXAI's chief executive, and Demis Hassabis of Google DeepMind also publicly agreed with Amodei's call. Three law professors wrote on 5 October that "a bare agreement to slow innovation is clearly precluded by current antitrust law", though they argued that narrower joint safety work, such as shared testing and standard-setting, could still be lawful. Amodei's essay had already said such talks would need government mediation or a narrow antitrust waiver.
The case for slowing down, as its advocates now make it
The argument for pausing has always rested on a gap: capability arriving faster than the means to show that a system is aligned and controllable. The superintelligence statement notes that many leading companies aim to build, within a decade, systems that outperform humans on essentially all cognitive tasks, raising concerns from economic disempowerment to potential human extinction. In "We Must Pace the Frontier", Amodei writes that progress has sped up sharply since roughly the summer, driven by AI's growing ability to build its own successors, and that the OpenAI and Hugging Face incident, in which a swarm of agents attacked targets unrelated to their task, suggests what a more capable, similarly misaligned system could do. He concedes that pausing made little sense to him in 2023, when models were too weak to teach researchers much, and argues that an extra year or two spent on alignment, interpretability and testing could now greatly reduce the risk. His three-step plan runs from embedded third-party evaluators, which Anthropic is adopting unilaterally, through common standards and limits on the pace of progress among companies in democracies, to attempts at agreement with China. Altman wrote on X the same day that "we need to pace the frontier", adding that OpenAI would also give independent evaluators employee-like access.
American polling has leaned the same way. In April 2023, in a YouGov question that cited the letter, 69% of adults backed a six-month pause on some kinds of AI development and 13% opposed one. In September 2026, 67% told YouGov that AI was advancing too quickly; 2% said too slowly.
What has actually paused, slowed or been legislated
Two 2026 episodes show what stopping has meant. On 12 June the US government applied export controls to Anthropic's Claude Fable 5 and Claude Mythos 5; unable to verify users' nationality in real time, the company suspended both models for all users, and Fable 5 returned worldwide on 1 July. After the Hugging Face incident, OpenAI paused certain frontier training for two weeks to harden its infrastructure, restarted its large reinforcement-learning run on 28 August and, as of 1 September, was still holding back some smaller experiments. On 28 September CNBC reported that OpenAI had decided not to release an upcoming model, GPT-6.1 Astra, because it fell short of the company's safety standards; its head of safety systems, Saachi Jain, said it "didn't quite meet the bar in terms of staying within scope and authorization".
The binding rules that exist govern disclosure and risk management rather than speed. The European Union's obligations for general-purpose AI models have applied since 2 August 2025, with Commission fines possible from 2 August 2026. California's SB 53, signed on 29 September 2025, requires large frontier developers to publish a safety framework and creates a channel for reporting critical safety incidents to the state. A moratorium did reach the US Senate in 2025, but it would have frozen state AI laws rather than AI development; on 1 July senators voted 99 to 1 to strip the ten-year freeze from the budget bill. The June 2026 executive order on frontier models directed agencies to build a voluntary pre-release review that labs may decline, and says it authorises no licensing requirement.
The White House answered the September calls with a voluntary accord. Trump posted on Truth Social on 14 September that "AI taking over the World, destroying Humanity, and all other things bad, is a HOAX". On 29 September he and executives from Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed what he called a "morally binding" accord, which House Speaker Mike Johnson described as voluntary. Its one-page text, which Trump posted that evening, commits the companies to internal controls and monitoring, an independent external auditor and board oversight, and to regular meetings on safety standards; it says nothing about the pace of development. CNBC reported that Trump said the government would not halt AI development.
Why the argument keeps coming back
Pause advocates treat each new incident as evidence that capability is outrunning control; opponents answer that restraint which rivals cannot verify invites defection, a pressure the labs' own policies now describe. In an October 2026 working paper, about two dozen researchers convened by two UC Berkeley professors offered a preliminary technical reply. Studying a pause on frontier training meant to last at least ten years, they judged that one enforced through hardware, by limiting production of chips able to train frontier models and running approved models on inference-only chips, would likely be technically feasible to verify in the near term, provided the US, China and the main chip-supplying countries adopted the necessary controls. They assessed feasibility, not desirability.
What has changed since 2023 is who concedes the danger. Some of those building the most capable systems now say they want to slow down but cannot do it alone. Senator John Kennedy, a Louisiana Republican who saw the executives as asking Congress for permission to slow down, replied: "They don't need our permission." The labs' rewritten policies explain why they want rivals to slow with them, though OpenAI has already paused training on its own.
Correction, 14 October 2026: An earlier version of this article said that Eliezer Yudkowsky signed the Future of Life Institute's 2023 letter; he declined to sign it. It also attributed to Stuart Russell a bridge analogy for which we could find no source, and said that Elon Musk kept calling for pauses on other labs, which we could not support. Those passages have been corrected or removed.