David Robinson, an OpenAI employee who wrote safety reports accompanying major model releases, has resigned and publicly criticized the culture shaping frontier AI development. The Verge reports that Robinson announced his departure this week and expanded on his concerns in an editorial published by The Atlantic.
Robinson's argument goes beyond adding another policy or compliance rule. According to The Verge, he describes an industry culture built around extreme confidence, relentless sprints and optimism that can minimize or overlook potential harms. He argues that this underlying mindset is the larger problem and that technical organizations developing powerful models need more humility.
His proposed standard comes from high-reliability industries rather than conventional software development. Robinson says frontier AI laboratories should operate more like nuclear power plants or busy airports, with multiple layers of redundancy and deliberate planning. The goal would be to prevent an ordinary human mistake from creating an opening for a catastrophic failure.
The Verge notes the understandable skepticism surrounding warnings from people who helped build the technology they later describe as dangerous. The article argues, however, that this history is not a reason to dismiss the substance of their concerns. Robinson's previous responsibility for model-release safety reporting gives his criticism direct relevance to how risks were communicated inside and outside OpenAI.
His resignation follows other departures from prominent AI companies. The Verge points to Jacob Coxon's exit from Anthropic and later warning about existential risk. It also names Robert O'Callahan, Bilal Chughtai and Josh Engels as departures from Google DeepMind, along with Joe Benton from Anthropic.
The report does not provide an OpenAI response to Robinson's claims, identify a specific safety failure that prompted his resignation or independently establish that the company's safeguards are inadequate. His comments are therefore an insider's diagnosis rather than proof of a particular incident. Even so, the departure adds another public challenge to the industry's ability to evaluate its own risks while racing to develop more capable systems.
Robinson's analogy implies a significant change in operating philosophy. Nuclear plants and airports are designed around the expectation that people and components will sometimes fail, so safety depends on overlapping protections rather than confidence in any single process. Whether frontier AI companies would accept the slower pace and added friction of that model remains uncertain.