Microsoft Deployed GPT-4 Before OpenAI Safety Board Finished Review. Now It Is Court Evidence.
When Rosie Campbell took the stand on May 7, she delivered a detailed inside account of OpenAI's safety governance that has been among the most specific testimony to surface in Elon Musk's lawsuit against the company. Campbell, who led OpenAI's AGI Readiness team until she left in late 2024, testified that Microsoft deployed a version of GPT-4 through Bing in India before OpenAI's own Deployment Safety Board finished evaluating the model. That board exists to catch what the company's own engineers might miss before a model reaches users at scale. Microsoft shipped first.
The bypass is not a theoretical concern. Campbell described it as the kind of precedent that compounds: if a safety process can be set aside once, it can be set aside again, and the people who bypassed it learn that it does not actually stop anything. "We want to have good safety processes in place we know are being followed reliably," Campbell testified, according to TechCrunch's reporting from the San Francisco courtroom. The gap between that aspiration and what she described from the inside is the distance between OpenAI's public safety commitments and its internal record.
The Deployment Safety Board, as Campbell described it, is a cross-functional group that reviews models for capability risks and known failure modes before a release is cleared. Former employees have described it as a gate that could be overridden when timelines compressed. Campbell's testimony is the first time a named former employee described a specific, dated instance of that circumvention under oath.
Tasha McCauley, an OpenAI board member who served during the period Campbell described, gave testimony that maps to the same pattern from the other side of the table. "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," McCauley testified, describing the board's access to information from Sam Altman. She said internal governance failures should drive stronger government regulation of artificial intelligence. "If it all comes down to one CEO making those decisions and we have the public good at stake, that is very suboptimal," she said.
David Schizer, an expert on nonprofit governance whom Musk's team called as a witness, testified that OpenAI must take its safety rules seriously and that what matters is whether those rules were followed in practice, not just whether they exist on paper. "What matters is the process issue," Schizer said, according to TechCrunch.
Campbell also testified to a cultural shift she witnessed from research organization to product company. "When I joined, it was very research-focused and common for people to talk about AGI and safety issues," she said. "Over time, it became more like a product-focused organization." Two safety-focused teams were dissolved during that transition: Campbell's AGI Readiness team and the Super Alignment team that had been led by Ilya Sutskever.
The lawsuit, filed in the Northern District of California, is Musk's attempt to force OpenAI to operate as a genuine nonprofit and to unwind what he argues is a betrayal of the company's founding agreement. The trial has surfaced internal messages, board deliberations, and former employee testimony that would otherwise remain private. What Campbell added on May 7 is a concrete, dateable instance of a safety gate being overridden, not in an internal Slack message but in sworn testimony by someone who was in the room when it happened.
There is a legitimate skeptical read of this testimony. Campbell left OpenAI over the grievances she described in court, and her account emerged in litigation brought by a co-founder who is now running a competing AI company. Campbell herself acknowledged under cross-examination that, in her view, OpenAI's safety approach remains superior to that of xAI, Musk's own AI venture. Schizer was retained by Musk's legal team. The picture these witnesses paint is not contested in a vacuum, but it is contested by interest and timing, and a reader should weigh that.
OpenAI has argued that its safety processes have continued to evolve and that its commitment to safe deployment is genuine. The company has pointed to its ongoing work in capability evaluation, red-teaming, and prepared-for-deployment protocols as evidence of a serious approach. The board McCauley described did not share that confidence at the time she served on it.
The deeper problem this testimony exposes is structural. There is no U.S. regulator with jurisdiction to audit OpenAI's Deployment Safety Board, no public filing requirement for internal safety reviews, and no independent body that regularly audits AI labs' internal governance processes. What Campbell and McCauley described in a courtroom is the closest thing the public has to an honest accounting of how those processes actually functioned during the period when OpenAI was making its most consequential decisions about what to release and when. That the accountability mechanism is a lawsuit filed by a billionaire competitor, rather than a regulatory proceeding or investigative report, tells its own story about where AI governance actually stands.
That creates a second-order problem. If safety-process failures become litigation evidence, AI companies have an incentive to reduce those failures to informal agreements and verbal approvals that leave no written record. The transparency that regulators and the public depend on is the first casualty of that shift. McCauley's own testimony acknowledged this: she said internal governance failures should drive stronger government regulation, meaning she was calling for the oversight mechanism that does not yet exist.
What to watch next: Judge Richard Seabright is expected to rule on Musk's central claim, that OpenAI's Microsoft partnership and product-first direction violated the nonprofit's founding charter, in the coming months. The testimony of Campbell and McCauley gives that claim a specific factual foundation inside the courtroom even if the legal outcome remains uncertain. Whether Congress uses any of this to build a regulatory framework for AI safety governance is the larger question that the trial itself cannot answer.