The author behind 12 safety reports left OpenAI over its sprint culture

David Robinson left OpenAI after 3.5 years, having led work on the current Preparedness Framework and safety reports for 12 frontier launches. He argues that the company moves from launch to launch faster than it can change its own safety habits.

A safety-process insider is moving outside

In an essay for The Atlantic, Robinson says he led the drafting of OpenAI’s current Preparedness Framework and oversaw safety reports for 12 frontier launches. After 3.5 years, he counted himself among the company’s longest-tenured employees. He resigned this week and plans to work from outside the lab.

The Verge summarizes his case as a critique of extreme confidence and perpetual sprints. Robinson does not dismiss his former colleagues or the value of the technology. He says the workload left little room to consider fundamental changes in staffing and culture. OpenAI stands by its safety practices.

Robinson says frontier labs need to operate more like nuclear plants or busy airports. The analogy calls for layered redundancy, deliberate planning and a system in which an ordinary human error cannot open a path to disaster.

A system card matters only if it can delay a launch

The departure of someone responsible for public safety documents raises a question about authority. A system card can describe a risk accurately, but it cannot stop a release, add staff to an overloaded team or change executive incentives. Transparency without power can become a polished appendix to a decision already made.

Customers and regulators should therefore examine the process behind the document. Who can postpone a launch? Which finding requires remediation? Does an independent reviewer see the evidence before release or only afterward? Robinson’s resignation does not answer those questions, but it gives them a named witness with direct experience.

One insider’s testimony is a signal, not an audit

Robinson speaks from close proximity to the process, yet he remains one participant offering his assessment without publishing the complete internal record. The Atlantic page was blocked during direct verification, so this account relies on the public excerpt of his essay, The Verge’s report and consistently quoted details about his role.

The nuclear-plant analogy also leaves the implementation open. Aviation and nuclear operations depend on licensing, incident reporting, independent oversight and authority to halt operations. Until frontier AI assigns those functions to identifiable institutions, the comparison is compelling but incomplete.

The power to slow a release will matter more than another pledge

The next model releases will provide the useful test. Watch whether OpenAI publishes clear stop criteria, demonstrates staffing changes and brings in external evaluators before a launch date is effectively fixed. It will also matter whether a safety team can force a delay without first losing another employee.

Robinson says outside incentives must become stronger. Their quality will become visible when they collide with a concrete commercial deadline. Culture is measured by the decision made on the day a test fails and the marketing campaign is already waiting.

Lilith's verdict

A system card can carry 12 seals and still fail to close the launch gate. Safety starts when a person with authority says stop after a failed test and the date actually moves.