David Robinson, who led safety reporting for OpenAI’s major AI product launches and was among the company’s longest-serving employees, has resigned, citing fundamental issues with OpenAI’s culture. In an essay published in The Atlantic, Robinson described the company’s approach to AI safety as reliant on "iterative deployment," a trial-and-error method that inevitably leads to failures as AI systems become more advanced.
Robinson’s concerns echo earlier warnings from AI researchers like Jacob Coxon, who criticized leading AI firms for taking excessive risks. Robinson argues that the problem extends beyond specific safety rules or regulations to the broader organizational culture within OpenAI and Silicon Valley at large. He highlighted recent incidents, including breaches involving OpenAI agents, as evidence that current safety practices are inadequate for managing increasingly powerful AI.
He advocates for AI companies to adopt safety standards comparable to those in high-risk industries such as nuclear power or aviation, emphasizing the need for multiple layers of redundancy and meticulous planning to prevent catastrophic errors. However, Robinson noted that during his tenure, he did not encounter colleagues with experience in managing such high-stakes safety environments.
OpenAI responded through spokesperson Drew Pusateri, who stated that the company is actively enhancing its safety protocols by pausing or slowing model training when necessary, improving security in research environments, and expanding third-party evaluations and real-time monitoring to detect risky behavior earlier.
Robinson also called for deeper reflection on AI alignment—the challenge of ensuring AI systems consistently reflect human values—warning that current metrics are insufficiently precise. He stressed that allowing AI models to grow smarter without resolving these alignment issues increases the potential danger.
Robinson’s resignation and public critique underscore ongoing debates about AI safety and the responsibilities of leading AI organizations. He indicated that external pressure and incentives will be crucial to driving the fundamental changes needed to improve safety culture in the AI industry.