OpenAI safety leader David Robinson says its culture is ‘broken’; company defends safeguards
David Robinson, a departing leader in OpenAI’s safety team, says the company’s culture is “broken” because rapid development and repeated product launches leave too little room for the level of care he believes advanced AI requires. Robinson led the writing of safety reports for major launches and helped draft OpenAI’s current Preparedness Framework. His assessment is a former employee’s characterization, not an established finding about the company. OpenAI says it is strengthening its safety practices and will pause training or hold back models when it needs to slow down.
Robinson set out his reasons for leaving in an essay published October 3, 2026. He said he had worked at OpenAI for three and a half years, led transparency work for its safety team and oversaw safety reports for 12 launches of advanced models. Those responsibilities gave him a role in explaining how the company evaluated risks, but they should not be confused with being OpenAI’s overall head of safety.
What Robinson meant by “broken”
Robinson’s criticism centers on how decisions are made under pressure. In his view, a workplace accustomed to ambitious goals and successive launch deadlines is too willing to identify problems through use and improve safeguards afterward. He argues that such an approach becomes less acceptable as AI systems grow more capable, because some failures might be difficult to contain or reverse.
He did not say his former colleagues were indifferent to safety. Robinson described them as capable people trying to make sound decisions, while arguing that the organization’s pace made it hard to step back and consider larger changes. His essay extends that concern beyond OpenAI to the wider industry. The claim that its culture is broken is therefore his judgment about priorities and working practices, rather than the result of a published investigation establishing a company-wide failure.
Robinson pointed to reported problems involving AI agents and safety controls as reasons for concern. He argued that frontier AI developers should draw more heavily on safety expertise from fields such as aviation and nuclear power, where systems are designed with multiple safeguards against human error. He also called for more research into whether increasingly capable models will behave safely when they are not being watched. Those are proposals and warnings about future risk, not evidence that the specific worst-case outcomes he describes have occurred.
How OpenAI responded
OpenAI spokesperson Drew Pusateri said the company is continuing to improve its safety and security measures. The company says it aims to prevent models from becoming more capable than it can safely manage and secure, and that it pauses training or withholds models when necessary. Pusateri also described work to strengthen research and testing environments, expand third-party evaluation and improve monitoring so concerning behavior can be detected earlier.
That response addresses the safety practices Robinson questions, though it does not establish that he is wrong about the company’s internal culture. Conversely, Robinson’s account offers an informed view of his experience but does not, by itself, demonstrate that OpenAI has abandoned safety work. The dispute is partly about how much confidence to place in safeguards that are being improved while the technology continues to advance.
Robinson said he plans to work outside OpenAI to help build stronger incentives for safety, though he had not settled on a specific next role. His departure brings a focused question into public view: whether the company’s ability to identify risks, delay work and change its practices can keep pace with the systems it is building. Robinson says the answer requires a deeper shift in culture; OpenAI says it is already taking steps to manage those risks.
