Ex-OpenAI Safety Writer Resigns, Says Industry Culture Is Broken

David Robinson, the person who authored the safety reports that accompanied OpenAI's major model releases, has left the company. This week he resigned and went public with his concerns in an editorial for The Atlantic.

Robinson's central argument is that the problem is not a gap in rules. In his telling, the culture of the AI industry is itself broken, and adding a few new regulations or tweaking how models are trained will not fix that. The issue runs deeper than any single policy.

His departure adds to a growing list of current and former insiders who have stepped forward to warn about the technology they helped build. It is easy to feel cynical about that pattern. These are, after all, people who took part in creating the systems they now describe as dangerous. But the criticism of the messenger does not answer the message, and Robinson's role gives his account some weight. Safety reports are the main public record of how a lab says it tested a model before shipping it. Someone who wrote them for every major release has a close view of how those assessments are made and how much influence they carry against commercial pressure.

The editorial arrives as AI companies race to ship ever more capable models on tight schedules, and as debate continues over whether voluntary safety commitments are enough. Robinson's framing suggests he does not think so, and that incentives inside the industry matter more than the paperwork produced around them.

OpenAI has not been the only company to face departures and public criticism from safety-minded staff. What stands out here is the specific job Robinson held. Rather than a peripheral critic, he was responsible for the documents meant to reassure the public that releases had been vetted.

It remains to be seen how OpenAI will respond, and whether Robinson's piece prompts any change in how the company or its rivals handle safety reporting. For now, his resignation is another data point for those arguing that the industry cannot be trusted to police itself.