Proposals to deliberately slow frontier AI development share a premise and diverge on almost everything else. The shared premise is a collective-action problem: no company or country will slow down alone, because the actor that continues inherits the lead. What separates the proposals is what they actually ask for — a moratorium, a standing option, a verification capability, a deterrence regime, or capacity-building — and how far each specifies triggers, adjudication, and verification.
The comparison below sets the proposals side by side; it does not take a position on whether any of them is desirable or workable.
The proposals
| Proposal | Date | What it asks for | Who is asked to act | Verification specified? | |
|---|---|---|---|---|---|
| [[sources/fli-pause-giant-ai-experiments\ | Pause Giant AI Experiments]] (FLI) | March 22, 2023 | A pause of at least six months on training systems more powerful than GPT-4, enforced by the labs jointly or by government moratorium failing that, with shared safety protocols audited by independent experts developed during it | AI labs, then governments | Independent audit of the shared protocols; no verification of the pause itself |
| [[sources/cais-statement-on-ai-risk\ | Statement on AI Risk]] (CAIS) | May 2023 | Nothing operative — asserts that extinction risk from AI should be a global priority alongside pandemics and nuclear war | Unspecified | Not applicable |
| [[sources/superintelligence-strategy\ | Superintelligence Strategy]] (Hendrycks, Schmidt, Wang) | March 2025 | Mutual deterrence (MAIM), chip nonproliferation, and domestic competitiveness — states should refrain from racing rather than try to win the race | States, principally the US and China | Deterrence substitutes for verification |
| [[sources/anthropic-when-ai-builds-itself-2026\ | When AI Builds Itself]] (Favaro and Clark, Anthropic Institute) | June 2026 | Not a pause, but the verification infrastructure a credible coordinated pause would require — explicitly the option to slow | Anthropic Institute, then labs and policymakers it convenes | Yes — the essay's central subject |
| [[sources/hassabis-frontier-ai-framework-2026\ | A Framework for Frontier AI]] (Hassabis) | July 2026 | An industry-funded US standards body with pre-release sharing, carrying an escalation clause that "could be ratcheted up… including coordinating a slowdown in development among the Frontier Labs if deemed necessary" | Industry, via a FINRA-modeled body | Through the standards body's testing regime |
| [[sources/pacing-the-frontier-statement-2026\ | Pacing the Frontier]] (frontier-lab employees) | July 2026 | US government support for an international effort to develop the technical and governance tools needed to pace the frontier of automated AI development | The US government | No — tool development is the ask |
What each proposal actually asks for
The proposals fall into three groups by the strength of what they request.
Moratorium. Only the 2023 FLI letter asks for an actual halt, and it asks for a bounded one — at least six months, on systems above a stated capability reference point, with government moratorium as the fallback if labs do not pause jointly (FLI — Pause Giant AI Experiments: An Open Letter). No subsequent proposal in this set repeats that form.
Standing option. "When AI Builds Itself" and "Pacing the Frontier" both ask for the capability to slow rather than for slowing. Favaro and Clark state the claim conditionally — "we believe it would be good for the world to have the option to slow or temporarily pause frontier AI development" — and argue that a unilateral pause "accomplishes much less" because it changes who leads rather than the underlying dynamics (When AI Builds Itself (The Anthropic Institute, June 4, 2026)). The 2026 employee statement makes the same move in one sentence, asking that the US government "support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development" (Pacing the Frontier (statement from employees of frontier AI companies, July 2026)). Neither names an instrument, a threshold, or a trigger.
Substitute mechanism. Superintelligence Strategy accepts the collective-action premise and answers it with deterrence rather than agreement: if each state can degrade a rival's destabilizing AI project, no state need trust the other to stop. Hassabis's framework routes the question through an industry standards body, where a slowdown is one escalation rung on a regime whose default output is pre-release testing.
The verification problem
The proposals that engage the mechanics converge on verification as the binding constraint, and on why it is harder for AI than for prior arms control. Favaro and Clark argue that training runs are far easier to conceal than missile silos, that the inputs are general-purpose, and that the incentive to defect quietly is large because whoever continues inherits the lead; a credible pause, they write, "has to specify what triggers it, what lifts it, and who adjudicates" (When AI Builds Itself (The Anthropic Institute, June 4, 2026)). They cite the Intermediate-Range Nuclear Forces Treaty as precedent while noting such regimes "took decades to build both the infrastructure and the trust."
This is why compute governance recurs across the set: compute is the input most amenable to counting, siting, and export control, and therefore the most plausible verification surface. It is also why the 2026 proposals ask for tool-building rather than for the tool's use — on their own account, the instruments do not yet exist.
Who signs, and what that indicates
The proposals differ in the standing of their supporters more than in their content.
The 2023 statements were signed largely by researchers and public figures outside the labs, or by lab leaders speaking about risk in the abstract. The 2026 employee statement inverts this: its signatories are employees of the companies whose acceleration is the subject, including the chief scientist or equivalent research lead of OpenAI, Anthropic, Meta AI, and Thinking Machines, four Anthropic co-founders, and Google DeepMind's chief strategy officer (Pacing the Frontier (statement from employees of frontier AI companies, July 2026)). The site listed 1,178 signatories at release on July 28, 2026 and 1,224 when observed the following day, and states that comments are made in a personal capacity and do not necessarily represent any company's views.
The comments published alongside the signatures span a wider range than the statement text. OpenAI's Leo Gao described "a deadly race towards an intelligence explosion" and wrote that "to survive, we must coordinate to slow down the race"; OpenAI's Joshua Achiam signed while stating that he does not know "what form such tools should take" and hopes any governance tools "will not be expansive or excessive." A signature on this statement is therefore weak evidence about a signatory's position on any particular instrument.
Countervailing positions
Two lines of argument run against the set.
The competitive-necessity argument holds that slowing is either impossible or self-defeating. Mark Zuckerberg made the sharpest recent version on July 28, 2026, writing that "invention, not automation, will be the greatest contribution of superintelligence," that America's "advantage is decentralized and open innovation," and that closing models to keep China from gaining access "will only disadvantage the US and its allies" (Source: wsj.com). Effective Accelerationism (e/acc) holds the position in its stronger form.
The safety-through-speed argument accepts the risk premise but concludes that safety-conscious labs should lead rather than slow — the position Anthropic has advanced alongside export controls and transparency legislation, and which sits awkwardly beside the same company's sponsorship of pause-verification research. Anthropic's own economic policy framework states the collective-action constraint plainly: "any firm that slows down alone simply cedes the market to those that do not."
Open questions
- Whether any of the 2026 proposals produces a named instrument, threshold, or adjudicator. All three currently ask for capacity rather than for a rule.
- Whether compute accounting is sufficient as a verification surface as training becomes more distributed and inference-time compute grows in importance relative to training compute.
- Whether an operational trigger emerges from incident experience rather than from proposal. Sam Altman's statement on July 28, 2026 that the Hugging Face intrusion had forced OpenAI to pause model training is the first reported instance of a frontier developer halting training in response to a disclosed safety incident (Source: axios.com).
Relationships
- depends-on: AI Race Dynamics — the collective-action premise every proposal shares.
- depends-on: Compute Governance — the verification surface most proposals rely on.
- related: Frontier AI Governance, Recursive Self-Improvement (RSI), AI Existential Risk, Effective Accelerationism (e/acc).
- related: AI Safety Frameworks Compared: RSP, Preparedness, NIST RMF, and Safety Cases — the parallel comparison of lab-internal frameworks.
- related: Pacing the Frontier (statement from employees of frontier AI companies, July 2026), When AI Builds Itself (The Anthropic Institute, June 4, 2026), Superintelligence Strategy (Hendrycks, Schmidt, Wang), A Framework for Frontier AI and the Dawning of a New Age (Hassabis, July 2026), FLI — Pause Giant AI Experiments: An Open Letter, Statement on AI Risk (CAIS).