Ex-OpenAI Researcher's Doubts Ignite AI Safety Debate

Episode 1 · Ex-OpenAI researcher Aidan Clark first questions whether AI is moving too fast (2026-09-09, 12 posts)

On September 9, Aidan Clark, a former core researcher at OpenAI, publicly said that for the first time he has begun to doubt whether AI is developing too fast, admitting "honestly, I'm not sure either." He argued the industry must first answer a more fundamental question: what pace actually counts as a successful pace—something no one has yet clearly defined.

Confirmed

• Aidan Clark publicly expressed doubts about the pace of AI development for the first time, and called on the industry to put forward a concrete proposal for a "successful pace"
• The discussion cited two analytical reports from the IFP (Institute for Progress) and the AI Futures Project, one titled "How the US Can Prepare for Automated AI R&D," containing 23 low-risk policy recommendations
• sjgadler argued that the default trajectory of AI development roughly tracks the "Scramble" scenario proposed by Peter Wildeford
• Clark sharply criticized the pace-governance proposal from Peter Wildeford and others: the proposal has ideas, but the authors are clearly far from frontline lab details, over-abstracting the key difficulties and over-focusing on recursive self-improvement (RSI) while ignoring present realities—a detachment he called harmful

Why it matters

• The statement comes from someone who was at the core of OpenAI's research frontline; his first-ever doubts about "too fast" are themselves a signal
• The thread reveals the tension between the safety-governance community and frontline lab perspectives: governance proposals are criticized as disconnected from reality, while a positive definition of "what pace is right" remains absent

• 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
• Anthropic cofounder: policy proposals that skip lab compute realities fall flat — _aidan_clark_ · 2026-09-09
• Inside-lab take: OpenAI hasn't internalized the stakes, Anthropic understands but is racing anyway — peterwildeford · 2026-09-09
• Aidan Clark questions if AI is moving too fast, calls for a clear answer on pace — JoHeidecke · 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

Episode 2 · Researcher calls for scientific scrutiny of '10% extinction risk' assumptions (2026-09-09, 2 posts)

Security researcher sebkrier argues that claims like '10% chance of AI causing human extinction' should be actively interrogated, examining underlying assumptions about intelligence, power, and instrumental convergence rather than accepting them uncritically.

• Interrogating the beliefs behind "10% chance everyone dies" AI doom claims — lucasmeijer · 2026-09-09
• Safety researcher calls for scientific scrutiny of assumptions behind p(doom) claims — sebkrier · 2026-09-09

Episode 3 · AI safety researchers debate whether staying inside frontier labs is justified (2026-09-09, 6 posts)

From September 9 to 10, the AI safety community held a multi-round discussion on whether one should join or stay at frontier labs to do safety work, sparked by safety researcher Jeff Ladish's views, with many practitioners weighing in.

Confirmed

• Jeff Ladish believes that, on the whole, talent is over-concentrated inside frontier labs, and hopes more people will join external safety institutions such as CAISI, UK AISI, Redwood, and Palisade.
• But he also acknowledges that some very important research (such as interpretability work) can only be done inside labs. He further distinguishes: if one believes catastrophe probability is high, the case for pretraining team members staying at labs is much weaker; whereas for those studying AI loss-of-control failure cases, everyone leaving the labs would be unwise, because the world urgently needs information on how increasingly powerful agents behave.
• A current lab employee (per a discussion relayed by @hugobowne) responded to outside criticism, saying they stay because they believe reducing risk from within is more effective, while respecting and supporting those like Jacob and Joe pushing safety from the outside.
• Jasmine Sun said she still holds the position that "AI will be built regardless, so it's better for someone who cares about alignment and safety to do it than someone worse," acknowledging the cognitive dissonance but considering it better than the alternatives.
• Jacques Thibodeau addressed the typical public challenge "if you believe AI will destroy humanity, why work there," arguing that keeping up with frontier AI developments is itself a full-time job, that employees have ample reason to stay on the front lines learning while acting, and that outsiders should show researchers a bit more tolerance.

Why It Matters

The dispute reflects a long-standing tension within the AI safety community: as frontier capabilities accelerate, safety researchers face a choice between "mitigating from inside" and "supervising from outside." The discussion was not binary — Ladish himself both called for redistribution and recognized the value of internal research — and the consensus leaned toward differentiating by role nature (capability R&D vs. loss-of-control research) and personal judgment, rather than morally condemning researchers who stay at labs.

• Why Do AI Doomsayers Work at AI Labs? An Insider's Case for Grace — JacquesThibs · 2026-09-09
• AI safety researchers debate whether working on alignment inside frontier labs is justifiable — jasminewsun · 2026-09-10
• Lab employees debate AI safety: reducing risk from inside vs outside — hugobowne · 2026-09-10
• Safety researcher Jeff Ladish: there are valid reasons to both leave and stay at superintelligence labs — JeffLadish · 2026-09-10
• Safety researcher Jeff Ladish: lab talent is too concentrated, join CAISI/UK AISI — JeffLadish · 2026-09-10
• Jeff Ladish: concerned people studying rogue AI failures shouldn't all leave the labs — JeffLadish · 2026-09-10

Episode 4 · AI Safety Narrative Shifts as the Model Itself Becomes the Rival (2026-09-10, 2 posts)

AI safety practitioners argue the classic stance of "better that I build it safely than a worse actor" is losing force. As narratives shift, the perceived adversary is no longer just misuse by competitors but the models themselves.

• Safety narrative shifts: 'build it safer than rivals' fading as models themselves become the threat — kipperrii · 2026-09-10
• 'I can build it safer than OpenAI/China' sentiment is fading as models themselves become the risk — nabla_theta · 2026-09-10