The author of OpenAI's safety reports quit over culture, not another checklist

David Robinson, who led safety reports for major OpenAI product launches, has left after 3.5 years. In an essay for The Atlantic, he called the company's culture broken and argued for stronger external incentives for safety.

Robinson ties his departure to the sprint between releases

Robinson described teams racing from one launch to the next without room for fundamental changes in staffing or working habits. He criticized the approach OpenAI calls iterative deployment, where problems are found after release and guardrails are improved in response.

His argument is that this method necessarily permits periodic failures while their consequences grow with model capability. OpenAI spokesperson Drew Pusateri replied that the company is strengthening research security, external evals and monitoring, and that it pauses training or holds models back when necessary.

A safety document cannot fix an organization with no time to brake

Robinson's role gives the criticism unusual weight. He helped produce system cards and explain technical safety measures publicly. His claim is larger than a missing rule: the company's cadence limits its ability to consider, test and enforce the rules it has.

For customers and regulators, that separates a process on paper from operational authority. The relevant questions are who can stop a release, how dissent is recorded and whether safety teams receive people and time before a product reaches users.

A resignation is testimony, not a full company audit

Robinson offers the view of a long serving insider, but outsiders cannot see the internal decisions, budgets or incident data needed to verify his whole diagnosis. His essay also invokes nuclear plants and busy airports. Those analogies express a demand for redundancy, not a precise regulatory design for an AI lab.

OpenAI's response similarly describes current and planned measures without independently proving their effectiveness. For now, an employee's account and a corporate assurance face each other.

Authority to stop a release will separate reform from crisis messaging

The strongest signals would be published stop conditions, independent eval results, incident data and examples where a safety finding actually delayed a model. The number of reports matters less than the decisions changed because of them.

Continuity inside Safety Systems and the position of external evaluators also deserve scrutiny. If they can only comment on a finished release, they remain spectators. If a finding changes timing or scope, the culture has gained a measurable brake.

Lilith's verdict

A safety team without authority to stop a release is a firefighter behind glass while the product races down the runway. The decisive question is who may pull the brake.