2 hours ago
3
Secret memos compiled by OpenAI co-founder and former chief scientist Ilya Sutskever allege chief executive Sam Altman engaged in a sustained pattern of lying to executives and board members, misrepresenting internal safety protocols, and manipulating colleagues. The allegations, which are denied by an OpenAI spokesperson, preceded Altman’s dramatic firing and reinstatement in late 2023.
The memos, which span approximately 70 pages of Slack messages, human resources documents and explanatory text, were disclosed in a New Yorker investigation published on Monday. They were compiled at the request of fellow board members and sent as disappearing messages to avoid detection. One memo opens with a list headed “Sam exhibits a consistent pattern of …” with the first item listed as “Lying”.
The revelations raise pointed questions about governance at what is now one of the world’s most valuable technology companies, reportedly preparing for an initial public offering at a potential valuation of $US1 trillion ($1.45 trillion).
The governance questions also carry direct implications for Australia, where ChatGPT has an estimated several million users and OpenAI’s models are embedded in Microsoft products used across the federal public service.
In a sworn deposition in a lawsuit lodged by Elon Musk against OpenAI, Altman and other OpenAI executives, seen by this masthead, Sutskever confirmed that characterisation was “clearly” his view at the time. He told lawyers he sent the memos as disappearing messages because he feared that if Altman “had become aware of these discussions, he would just find a way to make them disappear”. Asked what action he believed was appropriate, Sutskever gave a single-word answer: “Termination.” He testified he had been considering Altman’s removal for “at least a year” before it occurred.
The allegations are similar to those in a separate archive of more than 200 pages of contemporaneous notes assembled by Dario Amodei, the former OpenAI safety lead who left in 2020 to found rival firm Anthropic. Amodei’s notes, also previously undisclosed, purport to document what he describes as escalating deception over several years.
Board members independently raised concerns about Altman’s conduct, as detailed in the New Yorker report. Helen Toner, a board member and Australian AI policy expert, discovered that safety features Altman told the board had been approved for GPT-4 in December 2022 had not in fact received sign-off. Fellow board member Tasha McCauley was pulled aside after the same meeting by an employee who revealed that Microsoft had deployed an early ChatGPT version in India without completing a required safety review – something Altman had not disclosed during hours of board briefings.
In the memos, Sutskever also alleged that when OpenAI prepared to release GPT-4 Turbo, Altman reportedly told then-chief technology officer Mira Murati the model did not need safety approval, citing general counsel Jason Kwon. But when Murati checked with Kwon directly, he responded over Slack: “ugh ... confused where sam got that impression.”
An OpenAI spokeswoman dismissed the investigation as a rehash of old events. “Much of the piece revisits previously reported events through anonymous claims and selective anecdotes sourced from people with clear agendas,” the spokeswoman told this masthead, claiming that the most serious allegations stemmed from “Elon Musk’s smear campaign, which The New Yorker itself investigated and debunked”.
The company said it “strongly rejects any suggestion that OpenAI has traded safety for growth,” pointing to investments in youth safety protections and government partnerships. OpenAI said it now has more than 900 million weekly active users and is “a very different company: larger, more mature, more rigorously governed”.
Multiple senior Microsoft executives described the company’s relationship with Altman as fraught. Paul Graham, who recruited Altman to lead startup incubator Y Combinator, privately told colleagues that “Sam had been lying to us all the time,” according to the report. Former OpenAI chief technology officer Murati, who provided material for Sutskever’s memos, said: “We need institutions worthy of the power they wield.”
The revelations land as the Albanese government develops its mandatory guardrails for high-risk AI applications. The question of whether frontier AI companies can be trusted to self-report safety failures is central to that regulatory design.
Professor Toby Walsh, one of Australia’s leading AI researchers at UNSW Sydney, said OpenAI should be judged by the same standards now being applied to social media companies. The New Yorker investigation noted that OpenAI faces seven wrongful-death lawsuits alleging ChatGPT contributed to suicides and a murder.
“They ought to be put in the same bucket as social media companies like Meta that are increasingly becoming recognised as dangerous and careless,” Walsh told this masthead.
He said the prospect of a trillion-dollar IPO made self-regulation even less credible. “The commercial pressures are immense. The stakes are millions, perhaps billions of dollars for people like Sam Altman. Doing the right thing is hard in such circumstances. Surely we’ve learnt the lesson of social media that tech companies cannot be relied upon to comply.”
David Krueger, a prominent AI safety researcher who served as a research director on the founding team of the UK AI Security Institute, said the allegations suggested a pattern that should disqualify Altman from his role.
The Business Briefing newsletter delivers major stories, exclusive coverage and expert opinion. Sign up to get it every weekday morning.
