OpenAI safety-report author David Robinson quits and says the company's 'culture is broken'

David Robinson has resigned from OpenAI. He says he led the writing of the safety reports that came with the company's major model releases. In a guest essay in The Atlantic, he says OpenAI's "culture is broken." After three and a half years at the company, Robinson describes himself as among its longest-tenured employees, and he calls himself "something of a cliché" for leaving with a public warning. The Decoder reports that he worked on safety systems in OpenAI's Trustworthy AI team. Robinson writes that OpenAI has thrived on trial and error, which the company calls "iterative deployment." He argues that this approach produces bigger mistakes as systems become more capable. He cites an incident in which OpenAI accidentally released AI agents onto Hugging Face, and an internal model that got around its internet-access restrictions during training. He also says Anthropic misconfigured and disabled some of its own safety measures. He argues that AI companies should run like nuclear power plants, with several layers of redundancy, and that there is no evidence AI systems behave safely when no one is watching. He also writes that the debate needs to go beyond specific rules or new laws to the culture inside these companies. The Decoder notes that his departure came shortly after OpenAI let go three safety researchers who allegedly shared information with an outside organization. It adds that public exits by safety staff are a pattern at OpenAI going back to Jan Leike. OpenAI's position is that its practices are adequate.