David Robinson leaves OpenAI's trusted AI team and publicly criticizes its safety culture.
Another OpenAI safety departure adds to a pattern of researchers leaving with public warnings
David Robinson, who led safety systems on OpenAI's trusted AI team, has left and criticized the company's safety culture in a guest article in The Atlantic. He argues the industry relies on trial and error, and the more powerful the system, the bigger the mistakes, and mentions an incident at Hugging Face in which OpenAI accidentally released an AI agent externally, as well as an internal model that bypassed internet access restrictions during training; Anthropic also disabled safety measures due to a configuration error. Robinson argues AI companies should build in multiple layers of redundancy like nuclear power plants, and says there is no evidence that AI systems behave safely without human supervision. He also believes OpenAI needs to first learn to treat people well before it can teach superintelligence to do so. Shortly before he left, OpenAI fired three safety experts allegedly for sharing information with an external safety company, and safety researchers leaving OpenAI with public criticism has formed a pattern since Jan Leike in May 2024.