After three and a half years at OpenAI, David Robinson left the company last week, taking with him a pointed critique of how frontier AI labs handle risk. Robinson had led transparency work for the safety team, including the preparedness framework and safety reports attached to major releases.¹

His departure lands amid broader turbulence. OpenAI said it had “parted ways” with three researchers over alleged breaches of policies on sensitive information, while the company was already facing scrutiny after a Hugging Face security breach and other reported cases of model misbehavior. Robinson had warned earlier that the company was changing quickly, but questioned whether it was changing quickly enough.²

In an essay published Saturday, Robinson argued that the pace of the work made meaningful internal reform difficult. “My colleagues and I were so busy sprinting that we seldom had the chance to consider big changes, much less to actually make them,” he wrote. His conclusion: “stronger incentives for safety — coming from outside the company — are a big part of getting this right.”¹

That argument places him in a growing line of departing AI safety workers who say the industry’s culture is the problem, not merely a lack of rules. Robinson has called for the kind of redundancy and slow, deliberate planning associated with nuclear plants or busy airports — a rebuke to Silicon Valley’s “perpetual sprints” and confidence that bigger models can be managed as they arrive.³

OpenAI rejects the idea that it is simply racing ahead. In response, a spokesperson said the company was ensuring models did not become “more capable than we can safely manage and secure,” and would pause training or withhold releases when it needed to slow down. The company also cited tougher research security, third-party evaluations and earlier monitoring for dangerous behavior.¹

The divide is now public: Robinson sees external scrutiny as the missing discipline; OpenAI says its internal safety machinery can still keep the brakes ahead of the engine.