OpenAI safety report lead quits, calling the culture broken
David Robinson says trial-and-error deployment guarantees failures, and the company lacks people who know how to run safety-critical systems.
David Robinson led the team that wrote safety reports for OpenAI's major product launches. After three and a half years, he has resigned, saying the company's culture is broken. TechCrunch reported the departure on October 3, after Business Insider first reported it, and Robinson laid out his reasons in an essay published in The Atlantic.
The criticism lands because of who is making it. Robinson spent three and a half years at OpenAI and ran the team behind the safety reports that accompanied its biggest launches, the documents meant to show that a new system had been vetted before it reached the public.
His central charge is about method. According to TechCrunch, Robinson argued that OpenAI's practice of "iterative deployment," learning by trial and error, guarantees periodic failures, and that those failures grow more serious as the systems scale up.
As evidence, he pointed to breaches of Hugging Face systems by OpenAI agents. To him, that was less a one-off incident than a sign of how a deploy-first approach plays out once software can act on its own.
He offered a different model for how the work should run. Frontier AI companies, he wrote, should operate like nuclear power plants or busy airports, built with enough redundancy that no single mistake becomes a catastrophe. But he said he had never met a colleague at the company with experience making airplanes fly safely or keeping nuclear reactors running.
Robinson knew how his exit would read. He called himself "something of a cliché," an employee at a leading AI company issuing a dire warning on the way out the door. "An environment where things like this can happen is no place to grow artificial minds," he wrote.
OpenAI pushed back. Spokesperson Drew Pusateri told TechCrunch the company is improving its safety measures, pausing training when necessary, strengthening security and improving real-time monitoring to detect and respond to concerning behavior earlier in the training process.