David Robinson, a longtime member of OpenAI’s safety organization who said he led the writing of safety reports accompanying major product launches, has resigned from the company and publicly challenged the culture behind its approach to increasingly capable AI systems.
Robinson announced the departure in an essay published by The Atlantic under the headline “I Quit OpenAI Because Its Culture Is Broken.” His central argument is broader than a complaint about one model or one safety decision: he says frontier AI companies are moving too quickly and that OpenAI’s operating culture does not provide the level of care he believes more capable systems now require.
The distinction is important: Robinson’s departure and his work on safety reports are reported facts; the claim that OpenAI’s culture is “broken” is Robinson’s judgment about how the company operates.
The criticism centers on what Robinson describes as a trial-and-error approach to safety. OpenAI has long emphasized iterative deployment — releasing systems, observing how they behave in real use and improving safeguards as problems emerge. Robinson argues that the cost of that model changes as AI systems become more autonomous and capable, because a failure that can be patched after deployment may have more serious consequences than it did with earlier systems.
That is Robinson’s assessment, not an independently established finding that OpenAI’s culture is unsafe. But his proposed remedy makes the disagreement more concrete. He argues that frontier AI labs should borrow practices from high-reliability industries such as nuclear power and commercial aviation, where layered redundancy and careful planning are designed to prevent a single mistake from becoming a larger failure.
He also called for stronger safety science aimed at keeping future autonomous systems under control. The argument is that technical capability is advancing faster than the industry’s ability to demonstrate, with high confidence, how increasingly independent systems will behave in unfamiliar situations.
OpenAI disputes the implication that it is unwilling to slow down when safety requires it. In a statement reported by The Guardian, the company said it is continuing to strengthen its safety and security practices and that it pauses training or holds back models when it believes their capabilities cannot yet be safely managed or secured.
The distinction matters because the resignation is easy to reduce to a familiar Silicon Valley storyline: an employee leaves an AI lab and issues a warning. Robinson’s critique is more specific. He is arguing that safety should be treated as an operating discipline embedded in how a frontier lab works day to day, rather than primarily as a set of evaluations and guardrails that can be improved after problems appear.
OpenAI’s response, by contrast, emphasizes that its current process already includes the ability to stop, delay and strengthen safeguards. The unresolved question is therefore not whether safety work exists inside the company, but whether the surrounding development culture gives that work enough authority, time and redundancy as model capabilities increase.
That debate is likely to extend beyond Robinson’s departure. As AI systems take on longer-running tasks and more autonomous actions, the operational standards used by the companies building them will become a larger part of the safety argument — alongside model evaluations, security controls and regulation.
