AI Pause

An AI pause is a proposed halt to, or slowdown of, the development of the most capable AI systems until their safety can be shown. The term covers voluntary moratoriums, conditional pauses written into companies’ own safety policies and binding international agreements. It became a slogan with an open letter in March 2023, and whether a pause would reduce risk or merely shift it remains one of the central disputes of AI policy.

Where the term comes from

On March 22, 2023, the Future of Life Institute published “Pause Giant AI Experiments”. It called on all AI laboratories to pause for at least six months the training of systems more powerful than GPT-4, publicly and verifiably, and asked governments to impose a moratorium if they did not. The time was to be used for shared safety protocols audited by independent experts; the letter stressed that it did not mean a pause on AI development in general. Early signatories included Yoshua Bengio, Stuart Russell and Elon Musk, Reuters reported; the page now shows 31,810 signatures.

A week later Eliezer Yudkowsky argued in TIME that the letter asked too little. He called for an indefinite worldwide moratorium on large training runs, enforced by international agreement and, if necessary, by force. In May 2023 Joep Meindertsma founded the activist group PauseAI in Utrecht. No laboratory paused. Max Tegmark, the institute’s president, said six months later that he had never expected one; the aim had been to make the debate legitimate.

Later forms

• Conditional pauses. Anthropic’s Responsible Scaling Policy (September 19, 2023) implies that the company will temporarily pause training more powerful models if its scaling outstrips its ability to meet its own safety procedures.
• Prohibition. The institute’s Statement on Superintelligence (October 2025) calls for a prohibition on developing superintelligence until there is broad scientific consensus that it can be done safely and controllably, and strong public buy-in. Its signatories ranged from Geoffrey Hinton to Steve Bannon and Prince Harry, AP reported.
• A negotiated slowdown. In “AI 2040: Plan A” (July 9, 2026), the AI Futures Project sketches a verifiable agreement between the United States and China that would delay superintelligence from around 2030 to 2040. The authors call it “the least bad plan we currently know of”.

Arguments for

Supporters argue that laboratories are building systems faster than anyone can learn to understand or control them, and that competition locks them into a race none can leave alone. Only a coordinated, verifiable pause, on this view, changes the incentives. Public opinion leans toward caution: in a Quinnipiac poll published on September 30, 2026, 30 percent of Americans said AI companies should stop developing powerful systems until their safety can be evaluated and 47 percent said they should slow down; 5 percent wanted them to move faster.

Arguments against

Critics doubt that a pause is enforceable or wise. Bill Gates asked in April 2023 who exactly could be made to stop. Others warned that a pause would hand China an advantage, that it fixed on speculative threats rather than present harms, or that it hyped the products it warned against; Andrew Ng argued that governments should not pause technologies they do not understand. Accelerationists add that delay costs lives, and Vitalik Buterin warned that pause and licensing regimes could concentrate AI power in a few hands. The International AI Safety Report 2026 names the underlying “evidence dilemma”: acting too early risks ineffective rules, waiting risks leaving societies exposed.

Why it matters now

The question returned in autumn 2026, the season chronicled in Black October. On September 28 OpenAI confirmed that it had shelved the planned launch of GPT-6.1 Astra because the model fell short of its safety bar, Reuters reported – a withheld release, not a halt to development. The AI-made song Nothing Went Foom! counts the 2023 letter among failed alarms. Any position on a pause rests on two estimates: how likely catastrophe is, the subject of p(doom), and how much waiting costs.

Related

p(doom) · p(doom) Clock · Accelerationism · Containment · FOOM · Permanent Underclass · Black October · Nothing Went Foom! · Hit the Brakes — the satire at the top of the Doom 100

Sources: Future of Life Institute: Pause Giant AI Experiments (March 22, 2023) · Reuters via Malay Mail: Musk, experts urge pause on training AI systems more powerful than GPT-4 (March 29, 2023) · Eliezer Yudkowsky: Pausing AI Developments Isn’t Enough (TIME, March 29, 2023) · PauseAI: About · Semafor: Author of ‘pause AI’ letter reflects on its impact (September 22, 2023) · Anthropic: Responsible Scaling Policy (September 19, 2023) · Statement on Superintelligence · AP via News4JAX: Prince Harry, Meghan join call for ban on AI superintelligence (October 22, 2025) · AI Futures Project: AI 2040: Plan A (July 9, 2026) · AI 2040: Plan A (full scenario) · Quinnipiac University Poll (September 30, 2026) · Reuters via The Star: Bill Gates says calls to pause AI won’t ‘solve challenges’ (April 4, 2023) · Fortune: 5 reasons people are dunking on the pause letter (March 30, 2023) · Vitalik Buterin: My techno-optimism (November 27, 2023) · International AI Safety Report 2026 · Reuters via GMA: OpenAI shelves new AI model release over safety concerns (September 29, 2026)