Ex-OpenAI researcher Aidan Clark first doubts AI's pace, slams detached governance proposals

On September 9, Aidan Clark, a former OpenAI core researcher now at Anthropic (self-described as one of the "Ants"), said for the first time that he has begun to doubt whether AI is developing too fast, admitting "honestly, I'm not sure." He argued the industry must first answer a more fundamental question—what pace would count as success—and called for clear proposals within months, prioritizing human welfare. In the same threads he sharply criticized external AI-pace governance proposals, sparking ongoing debate between the safety-governance community and frontier-lab insiders.

Confirmed

• Clark publicly expressed doubts about AI's pace for the first time and called for a positive definition of a "successful pace"
• The discussion cited two analyses from IFP (Institute for Progress) and AI Futures Project, one titled on preparing the US for automated AI R&D, with 23 low-risk policy recommendations
• sjgadler argued the default trajectory roughly matches Peter Wildeford's "Scramble" scenario
• Clark criticized governance proposals as well-intentioned but written by people far from frontier-lab details: over-abstracting key difficulties, over-focusing on recursive self-improvement (RSI), and ignoring present realities—a detachment he called harmful
• His litmus test: any pace-governance proposal that doesn't frequently mention Anthropic and OpenAI is likely at the wrong level of detail, since those two labs determine actual frontier progress; he offered to help fill in details
• He noted outsiders generally lack accurate knowledge of labs' compute and operations, making "reach into the box" style interventions unsound
• Responding to "if you believe the risk, why build AI," Clark said: if you think AI has 10% probability of exterminating humanity, the only defense for continuing is treating lab staff as "bad people." A widely shared tweet added that many at OpenAI haven't internalized civilizational risk, while Anthropic understands it but feels trapped in a race, believing no other lab will act responsibly
• On risk communication, BronsonSchoen argued "we're the responsible ones so we must continue" framing is also problematic and doom-laden messaging devolves into companies blaming each other; ericlim said the community underestimates AI's magnitude and lacks a "this helps now" narrative—positive framing spreads better than doomerism

Why it matters

• The statement comes from someone embedded in frontier research at both OpenAI and Anthropic; his first doubts about pace are themselves a signal
• The threads lay bare the tension between governance circles and frontier-lab insiders: proposals criticized as detached, the "why build if you believe risk" question answered firsthand, while a positive definition of the right pace remains missing

2026-09-09 ~ 2026-09-09 · 12 related posts

• [source] Former OpenAI researcher Aidan Clark: for the first time I'm asking if AI is moving too fast — _aidan_clark_ · 2026-09-09
• IFP proposes 23 low-regret policies for the US to prepare for automated AI R&D — sjgadler · 2026-09-09
• Default AI scenario is Peter Wildeford's 'Scramble', argue AI safety discussants — sjgadler · 2026-09-09
• Frontier Lab Researcher: AI Safety Proposals Over-Focus on RSI, Miss Real-World Details — _aidan_clark_ · 2026-09-09
• Frontier lab insider slams AI pacing proposals: detached from reality, over-focused on RSI — sjgadler · 2026-09-09
• Any serious AI pacing proposal must name Anthropic and OpenAI, argues former DeepMind researcher — _aidan_clark_ · 2026-09-09
• [source] Anthropic cofounder: policy proposals that skip lab compute realities fall flat — _aidan_clark_ · 2026-09-09
• [source] Inside-lab take: OpenAI hasn't internalized the stakes, Anthropic understands but is racing anyway — peterwildeford · 2026-09-09
• Anthropic researcher on 10% extinction risk: 'we're not awful people, we want to do good' — _aidan_clark_ · 2026-09-09
• AI safety debate: is doom-flavored risk communication from OpenAI and Anthropic actually working? — BronsonSchoen · 2026-09-09
• ericlim: positive AI stories beat doomerism as communication strategy — ericlim · 2026-09-09

1 near-duplicate retellings: JoHeidecke