Former OpenAI safety lead says trial and error is no longer enough

David Robinson, who led safety-transparency work at OpenAI and helped produce reports for major model launches, resigned after three and a half years and said the company’s culture is broken. His objection is to “iterative deployment”: releasing systems, finding weaknesses in use or testing, then strengthening protections. That approach can make sense for ordinary software, he argues, but it also accepts that failures will happen—and the cost of a failure rises as AI systems become more capable. Robinson wants leading AI labs to operate more like nuclear plants or airports, with redundant safeguards and slower planning so one mistake cannot cascade; OpenAI said it continues to improve safety measures and can pause training or withhold models when needed.