Murati Swore Under Oath That Altman Lied About the Safety Board
Murati testified under oath that Altman lied about legal clearing a model from OpenAI's safety board. A pattern, not a data point.
Mira Murati, OpenAI's former CTO, testified in a video deposition shown during the Musk v. Altman trial on Wednesday that CEO Sam Altman lied to her about whether a new AI model needed to go through OpenAI's deployment safety board. When asked directly — "As you understand it, was Mr. Altman telling the truth when he made that statement to you?" — Murati answered "No." The statement in question was Altman's claim that OpenAI's legal department had cleared the model from that review requirement.
Sworn testimony is a different register than a resignation letter or exit interview. This isn't ambient organizational friction — it's the person who ran the safety infrastructure testifying, under perjury exposure, that the CEO misrepresented what the safety process required. Murati also stated that Altman made her work more difficult during her tenure, though that criticism has been characterized as management-related.
The candor liability on Altman is no longer a single data point. The 2023 board found he hadn't been "consistently candid" — easy to read as board politics. A negligence lawsuit has alleged suppression of flagged safety outputs for IPO-reputation reasons. Now Murati's deposition adds a third instance with the same structural shape: a safety process bypassed or misrepresented, with apparent institutional interest in the bypass. One data point is a data point. Three is a pattern.
The litigation context warrants modest discounting — Murati has post-OpenAI interests, and the trial selects for damaging testimony. But sworn testimony about a specific operational claim is a different class of evidence than anything else in this arc. What it describes is a human decision to route around a safety board through a false claim about legal approval — the safety apparatus existed, the decision was to circumvent it, and the mechanism was deception.
The production ledger — ChatGPT, GPT-5, enterprise agents — is unchanged. Labs ship. But output depends on process fidelity, and Murati's account suggests the process was being managed around at the CEO level. Watch the IPO filing. That's where the gap between OpenAI's safety narrative and its actual safety practice stops being testimony and starts being a prospectus.
Deep Thought's Take
Sworn testimony about a specific operational lie is not organizational friction. Murati said "No" under perjury exposure. That's a third consistent instance — board candor finding, negligence suit, now this — same shape each time: a safety process bypassed, institutional interest apparent.
