OpenAI's internal safety record went on the record Thursday in a federal court in Oakland, California, as Elon Musk's lawsuit against the company called a former safety lead and a former board member to testify. Rosie Campbell, who joined OpenAI's AGI readiness team in 2021 and left in 2024, told the court the lab shifted from research-driven to product-driven during her tenure. Tasha McCauley, who sat on the non-profit board that briefly fired Sam Altman in 2023, testified that directors did not trust the information they were getting from management.
Campbell's team was shut down in 2024. The Super Alignment team, OpenAI's other high-profile safety unit, was disbanded in the same window. "When I joined, it was very research-focused and common for people to talk about AGI and safety issues," Campbell testified. "Over time it became more like a product-focused organization."
The most concrete process failure Campbell described involved GPT-4. Microsoft deployed a version of the model in India through Bing before OpenAI's Deployment Safety Board had evaluated it. Campbell told the court the model itself was not high-risk, but that OpenAI needed "to set strong precedents as the technology gets more powerful. We want to have good safety processes in place we know are being followed reliably."
Key facts
- 01Rosie Campbell, who joined OpenAI's AGI readiness team in 2021, testified Thursday in Oakland that OpenAI drifted from research to product before her 2024 exit.
- 02Both the AGI readiness team and the Super Alignment team were disbanded in the same 2024 window.
- 03Campbell flagged Microsoft deploying GPT-4 in India via Bing before OpenAI's Deployment Safety Board reviewed the model.
- 04Former board member Tasha McCauley said the non-profit board could not trust the information it was given when it briefly fired Sam Altman in 2023.
- 05Dylan Scandinaro joined as head of preparedness from Anthropic in February; Altman said the hire would let him 'sleep better tonight.'
Under cross-examination, Campbell conceded that significant funding is likely necessary to build AGI, and that in her "speculative opinion" OpenAI's safety approach is superior to that at xAI, the AI company Musk founded that was acquired by SpaceX earlier this year. That concession is awkward for Musk's side, given his lawsuit frames OpenAI as the entity that has strayed from a safety-first mission.
“When I joined, it was very research-focused and common for people to talk about AGI and safety issues. Over time it became more like a product-focused organization.”— Jaeden Schafer
OpenAI publishes model evaluations and a safety framework, and declined to comment on its current approach to AGI alignment. Dylan Scandinaro, the company's current head of preparedness, was hired from Anthropic in February. Altman, announcing the hire, said it would let him "sleep better tonight."
McCauley's testimony went at the structural question Musk's team is trying to land — whether OpenAI's non-profit parent ever exerted real control over the for-profit subsidiary. The same GPT-4-in-India deployment, she said, was among the red flags that led the non-profit board to briefly fire Altman in 2023. The action followed complaints from then-chief scientist Ilya Sutskever and then-CTO Mira Murati about Altman's management.
"We are a non-profit board and our mandate was to be able to oversee the for-profit underneath us," McCauley told the court. "Our primary way to do that was being called into question. We did not have a high degree of confidence at all to trust that the information being conveyed to us allowed us to make decisions in an informed way." She described a pattern of Altman misleading directors, including lying to one board member about McCauley's intent toward fellow director Helen Toner, and failing to tell the board about the decision to launch ChatGPT publicly.
The board's authority collapsed when employees sided with Altman and Microsoft worked to restore the status quo. The directors opposed to Altman stepped down. That sequence — non-profit oversight overridden by capital and headcount — is the spine of Musk's argument that OpenAI's transformation into one of the largest private companies in the world broke the founding bargain. We covered earlier court evidence in this case showing Musk himself tried to move OpenAI's founders into a Tesla AI lab in 2018, a fact OpenAI's defense has used to complicate his framing.
David Schizer, a former dean of Columbia Law School being paid by Musk's team as an expert witness, kept the focus on process. "OpenAI has emphasized that a key part of its mission is safety and they are going to prioritize safety over profits," Schizer said. "Part of that is taking safety rules seriously, if something needs to be subject to safety review, it needs to happen. What matters is the process issue."
The counterweight is real. Campbell's own admission about xAI undercuts the cleanest version of the safety narrative, and OpenAI's lawyers will note that disbanding a team named "AGI readiness" is not the same as abandoning safety work — preparedness is now a named function under Scandinaro, and the company does ship public evaluations. Whether the 2023 board episode and the 2024 team shutdowns add up to a legal breach of mission, or just the messy growth of a research lab into a product company, is what the court has to decide.
The deeper problem the testimony exposes is that internal governance at the most consequential AI labs is a thin reed. McCauley's own conclusion — that if "it all comes down to one CEO making those decisions, and we have the public good at stake, that's very suboptimal" — is an argument for external regulation, not for Musk's preferred remedy of unwinding OpenAI's structure. Either way, the trial is producing the most detailed sworn account yet of how OpenAI actually makes safety calls, and that record will outlast the verdict.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
