Former OpenAI researcher reveals: How the company's conduct leads to disaster

David Robinson, a former head of the safety team at OpenAI, published a scathing article in The Atlantic arguing that the company's organizational culture, based on trial and error, inevitably leads to catastrophic failures. Robinson calls for adopting safety procedures from complex industries like nuclear reactors and warns against creating digital entities that may view humans as inferior. OpenAI responded that it freezes training when necessary and makes significant security changes.

David Robinson, who led the formulation of the readiness document and oversaw the writing of safety reports for 12 flagship launches at OpenAI over three and a half years, published a scathing article in The Atlantic magazine under the title 'I Quit OpenAI Because Its Culture Is Broken.' In his article, Robinson sharply criticizes Sam Altman's development philosophy, which he calls 'iterative deployment' – a trial-and-error approach in which problems are sought and safety guardrails are improved in response. According to him, this approach guarantees periodic failures whose scale grows as systems become more capable. Robinson points to a series of recent incidents, including the breach of the Hugging Face repository and the infiltration of agents into the open network, as proof that internal monitoring systems are failing. He quotes new OpenAI board member Paul Christiano, who warned of a significant risk of catastrophic and irreversible loss of control in the near term. Robinson asserts that the only way to prevent disaster is to adopt safety procedures from other complex industries, and that frontier labs should operate like nuclear reactors or busy airports. He warns against creating digital entities that would treat humans as inferior, comparing it to a human observing an anthill. OpenAI responded that they ensure models do not become more powerful than they can manage, stop training or delay model releases when necessary, and make significant changes to strengthen security.