The woman who built GPT-4 just told a court, under oath, that the CEO lied to her about safety protocols.
The Summary
- Mira Murati testified that Sam Altman falsely told her OpenAI's legal department had cleared a new AI model to skip the deployment safety board, a statement she confirmed under oath was untrue.
- The deposition surfaced in the Musk v. Altman trial, where Barry Diller defended Altman while simultaneously arguing that trust becomes "irrelevant" as we approach AGI.
- OpenAI's former CTO, who left in September 2024, stated Altman made her work "more difficult" during her tenure.
- The testimony exposes the central tension in AI governance: whether individual leadership integrity matters when building systems designed to exceed human control.
The Signal
Murati's testimony is damning not because it reveals personal animosity, but because it documents a specific lie about a specific safety process. When asked directly if Altman told the truth about legal's approval to bypass the safety board, she answered with one word: "No." This isn't office politics. This is the technical leader responsible for GPT-4's development saying the CEO misrepresented internal safety protocols.
The timing matters. Murati left OpenAI in September 2024, part of a broader exodus of safety-focused leadership. Her departure followed the dissolution of OpenAI's Superalignment team and the exits of Jan Leike and Ilya Sutskever. Now we know what was happening behind closed doors: disagreements over whether new models needed formal safety review.
"The woman who understood the models better than almost anyone else couldn't trust what her CEO told her about deploying them."
Enter Barry Diller with a different kind of bombshell. The IAC chairman told the court he trusts Altman personally, then pivoted to argue trust itself becomes meaningless as AGI approaches. His logic: once you build a system smarter than its creators, personal integrity of the leadership team is beside the point. You need structural guardrails, not trustworthy executives.
Key contradictions in the courtroom:
- Diller defends Altman's character while arguing character doesn't matter
- Murati says Altman lied about safety while Diller says we need safety beyond trust
- The defense and prosecution accidentally agree: trust isn't enough
This creates a strange legal moment where both sides are making the same structural argument while disagreeing on the facts. Diller's testimony, meant to defend Altman, actually validates the case for binding external oversight. If trust is irrelevant when building AGI, then Murati's testimony about broken trust is almost beside the point. What matters is whether OpenAI had robust safety processes, and whether Altman circumvented them.
The practical question: what happens when the CTO can't verify what the CEO says about safety protocols? Murati's statement that Altman made her work "more difficult" reads differently in this context. She wasn't complaining about management style. She was describing the experience of trying to enforce safety standards when leadership wanted to move faster.
The Implication
This testimony will shape how every AI lab structures internal safety processes. If OpenAI's CTO couldn't get straight answers about whether a model needed safety review, no amount of blog posts about "responsible AI" means anything. The question becomes: who has veto power over deployment, and can that power be overridden by the CEO?
Watch for two outcomes. First, investor pressure for independent safety boards with actual authority, not advisory roles. Second, talent flow. If you're a top AI researcher trying to decide between Anthropic, Google DeepMind, or OpenAI, Murati's testimony is a data point about internal culture that no recruiting pitch can overcome.
The Diller angle points to the endgame: regulation that doesn't depend on trusting anyone. If AGI makes trust irrelevant, then governance structures need to work even when leadership lies, makes mistakes, or gets replaced. That's not a dig at Altman. It's physics. You don't build containment vessels for nuclear reactors based on trusting the plant manager. You build them because physics doesn't care about trust.