← Back to feed

David Robinson, who led the writing of safety reports accompanying OpenAI’s major product launches, has resigned — publishing a public essay in The Atlantic calling the company’s internal culture “broken” and its approach to AI development insufficiently cautious.

Robinson acknowledged the optics upfront. By his own description, he is “something of a cliché”: an AI safety insider at a leading lab who departs and issues warnings on the way out. He made the choice anyway, and chose the highest-circulation venue available.

What He Said

The structural complaint in Robinson’s essay is pointed: the companies developing frontier models simultaneously tell the world that their systems are becoming more capable at extraordinary speed while warning that those same capabilities could create risks that are difficult to predict. His argument is that releasing more capable systems is not an adequate response to that tension.

The piece, published this week in The Atlantic, also addresses OpenAI’s internal culture directly. Robinson’s description of it as “broken” follows a long line of departures from the company’s safety-focused staff over the past two years, including co-founder Ilya Sutskever and Superalignment co-lead Jan Leike.

Separately, a Ctech headline citing his essay carries the phrase “the time for trial and error is over” — framing the core argument as one of urgency, not pessimism.

Context

Robinson is not the first to leave OpenAI citing safety concerns. His role was specific: he worked on the safety systems team for roughly three and a half years and oversaw safety reports and system cards for twelve frontier model launches, including drafting the Preparedness Framework. That proximity makes his account of internal culture harder to dismiss as observer bias.

OpenAI issued a public statement in response to the essay, saying the company continues to strengthen safety practices and pauses training or holds back models when necessary.

The timing is notable. OpenAI is in the middle of deploying autonomous agents at scale — Dots, its always-on assistant built around GPT-6 Astra, launched last week in a phased rollout for Pro and Business Premium subscribers with an Enterprise beta — and the regulatory environment is tightening. The FTC has opened a probe into AI product risk at both OpenAI and Anthropic. The UK and EU have each advanced their own oversight frameworks.

Pattern

A safety-focused departure from a frontier lab is, by now, a recognizable genre. More senior figures — including co-founder and Chief Scientist Ilya Sutskever and Superalignment co-lead Jan Leike — have left OpenAI in the past two years. What makes Robinson’s case different is his specific role: he did not work in general research or policy at a remove from deployment decisions. He wrote the safety documentation that goes out alongside the products. His claim is about the mechanism for resolving safety disagreements, not any individual call.

The argument OpenAI will likely make in response — to whatever degree it engages — is that it publishes safety cards, system cards, and evaluations more comprehensively than any other lab, and that Robinson himself contributed to that infrastructure. Robinson’s counter, implicit in the essay structure, is that publishing documentation is not the same as allowing it to change decisions.

Neither claim is unfalsifiable. What is falsifiable, eventually, is whether the systems OpenAI ships cause the harms that safety departures have warned about. So far they have not. Whether “so far” holds as agentic deployment deepens is the actual question.