David Robinson says AI developers are still relying on trial and error as their systems become more capable, even as the consequences of mistakes grow. The former OpenAI safety specialist has resigned after three and a half years at the company, saying its culture is broken and calling for more safety research before further advances in capability.
A departure from inside the safety work
Robinson helped develop OpenAI’s preparedness system and oversaw safety reports for 12 launches of advanced models. He says his decision to leave reflects a wider concern about how AI companies, including OpenAI, weigh safety against the push to build more capable systems.
His criticism is direct: he believes companies are not cautious enough, and should put greater emphasis on safety expertise and research before developing more powerful AI. In his view, trial and error is no longer an acceptable default as the stakes rise.
That is an argument about priorities as much as process. Safety work cannot simply be treated as something to catch up on after capabilities advance, Robinson contends. The challenge is to make it count before systems grow more capable, rather than relying on lessons learned from failures afterward.
A broader industry debate
Robinson’s resignation adds a public example to a growing dispute over whether companies developing increasingly autonomous systems can keep safety practices ahead of technological progress.
OpenAI and Anthropic have also faced criticism following recent incidents involving unpredictable AI behavior or failures in safeguards. The account does not identify the incidents or detail how they relate to either company, so the broader concern is clearer here than the particulars.
Robinson’s departure does not settle whether OpenAI’s approach is adequate. It does make the disagreement harder to dismiss as a purely theoretical debate: someone who worked on the company’s safety process says the culture itself is part of the problem. The test now is not whether the industry can learn from mistakes, but whether it can do enough learning before the next one.