Ex-OpenAI researcher tells MPs AI risk tolerance is 'catastrophically high'
Daniel Kokotajlo, who quit OpenAI over safety concerns, told MPs on Tuesday that the industry's risk tolerance is 'catastrophically high'. Two more ex-lab researchers gave similar warnings as MPs test whether the firms can regulate themselves.
The Board Room · Leadership desk
What happened
- Kokotajlo said he resigned after losing confidence that OpenAI would act responsibly as it pursued superintelligence, meaning systems that outperform humans across virtually all tasks.
- He alleged that staff faced pressure to play down their concerns, and said some internal research and forecasts had been kept from publication.
- Steven Adler, another former OpenAI researcher, told MPs that people inside the company had pushed back against calling its technology dangerous.
- Geoffrey Irving, who has worked at OpenAI, Anthropic and the UK's AI Security Institute, said capability advances were far outstripping advances in safety research.
Why it matters
- What doesn't add up contradiction Both firms have recently called for a slowdown and Anthropic's IPO filings reportedly cite existential risks, yet Kokotajlo describes pressure to soften public statements; the source does not date his departure, so these may be different periods.
- Who has to decide decision Parliament is examining whether the firms can regulate themselves, and the report of the hearing includes no draft rule or timetable for an outside limit.
- What it limits constraint The reported testimony speaks mainly to OpenAI: all three witnesses quoted worked there, so it says little about practice at Meta or Google DeepMind.
Adler's evidence bears most directly on self-regulation, because it describes an incentive. He warned that competition was pushing companies to justify ever riskier decisions by pointing to what their rivals were doing [8]. If that is accurate, a lab that holds a lower line alone gives ground to rivals who do not. We think that is the strongest argument in the hearing for an outside limit, and Adler urged governments to impose clearer limits [9]. In his words: "What I hope governments will do is they will help set the ceiling essentially on how much risk these companies can impose" [10].
A ceiling set by government needs government to measure the risk first. Washington restored the UK's access to advanced American AI models for pre-release safety testing after a temporary restriction, and City AM reported that the restriction raised questions about Britain's ability to scrutinise the technology independently [17]. Adler's request therefore depends on an access arrangement that Washington controls and has already restricted once.
Irving's figure needs the most care. Pressed on whether he still held to his estimate that advanced AI could carry a 50 per cent chance of destroying humanity, he said a "coin flip is roughly the right model to have" [6]. City AM noted that the figure reflects his personal assessment of a hypothetical future risk and is not an established scientific probability [7]. It shows how seriously one informed witness takes the problem, but gives a regulator no number to set a limit against.
The testimony varies in how far it can be checked. Adler left OpenAI in late 2024 and conceded that the company's internal culture could be different now, so his account of resistance to the word "dangerous" may not describe the company today [4]. Kokotajlo's account of blocked publication is different: it points to internal research and forecasts inside OpenAI, so the company can confirm or contest it [16]. His judgement that industry risk tolerance is catastrophically high cannot be tested the same way. City AM said it had contacted OpenAI and Anthropic for comment; its report does not include a response [11].
What to watch
- What the OpenAI, Anthropic, Meta and Google DeepMind executives told MPs later on Tuesday, and whether any of them disputes Kokotajlo's account of withheld research.
- Whether the committee's follow-up turns Adler's hoped-for ceiling into a specific limit with a timetable.
- Whether the UK's pre-release access to American models stays in place, since any ceiling would depend on testing them.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption
- Insufficient
- Hype gap+20
- Incentives
- Insufficient
- Confidence50
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Daniel Kokotajlo, a former OpenAI researcher who quit over safety concerns, told the Business, Innovation, Science and Trade Committee on Tuesday that the industry was developing increasingly powerful AI models without knowing how to control them.
- [2]
Kokotajlo told MPs: "I would say that the risk tolerance across the industry is catastrophically high".
- [3]
Steven Adler, another former OpenAI researcher, told MPs there had been resistance within the company to describing its technology as dangerous.
- [4]
Adler left OpenAI in late 2024 and acknowledged that its internal culture may have changed since his departure.
- [5]
Geoffrey Irving, a former researcher at OpenAI, Anthropic and Britain's AI Security Institute, warned MPs that progress in AI capabilities was far outstripping advances in safety research.
- [6]
Asked whether he stood by his assessment that advanced AI could have a 50 per cent chance of destroying humanity, Irving said a "coin flip is roughly the right model to have".
- [7]
City AM noted that Irving's figure reflects his personal assessment of a hypothetical future risk, rather than an established scientific probability.
- [8]
Adler warned that competition was encouraging companies to justify increasingly dangerous decisions based on what their rivals were doing.
- [9]
Adler urged governments to impose clearer limits on AI developers.
- [10]
Adler said: "What I hope governments will do is they will help set the ceiling essentially on how much risk these companies can impose".
- [11]
OpenAI and Anthropic have been contacted for comment.
- [12]
Parliament is examining whether tech giants can be trusted to regulate themselves.
- [13]
Executives from OpenAI, Anthropic, Meta and Google DeepMind were due to face MPs later on Tuesday, followed by senior government officials.
- [14]
All three witnesses quoted by City AM (Kokotajlo, Adler, Irving) had worked at OpenAI.
- [15]
Kokotajlo said he resigned after losing confidence in OpenAI's willingness to act responsibly as it pursued superintelligence, meaning systems capable of outperforming humans across virtually all tasks.
- [16]
Kokotajlo alleged that staff had faced pressure to play down concerns about the technology, and said some internal research and forecasts were stopped from being published.
- [17]
Washington restored the UK's access to advanced American AI models for pre-release safety testing, following a temporary restriction that raised questions about Britain's ability to scrutinise the technology independently.
- [18]
OpenAI and Anthropic have both recently called for a slowdown in the development of the most advanced AI models to enable safety measures to keep pace.
- [19]
In filings for its upcoming IPO, Anthropic reportedly said its technology could pose "existential risks to humanity".
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- cityam.comAI firms taking ‘catastrophically high’ risks, ex-OpenAI researcher warns MPs
1 article · October 13, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- UK parliamentary AI scrutinyFollow
- AI safetyFollow
- AI RegulationFollow
Entities
- Daniel KokotajloFollow
- Steven AdlerFollow
- Geoffrey IrvingFollow
- OpenAIFollow
- AnthropicFollow
- Google DeepMindFollow
- MetaFollow
- UK AI Security InstituteFollow
- Business, Innovation, Science and Trade CommitteeFollow