Departing OpenAI safety staffer says era of trial and error is over

The latest:

A former OpenAI safety employee argued that the company’s speed-driven culture is raising the odds of serious failures, in an essay published by the Atlantic on Saturday. David Robinson, who said he spent three and a half years at the company, wrote that advanced AI needs safeguards closer to those governing nuclear power and aviation. OpenAI said it pauses training or holds back models when it needs to slow down.

Details:

• The essay: Robinson’s piece, titled I Quit OpenAI Because Its Culture Is Broken, was published by the Atlantic on Saturday, Reuters reported. He argued that AI companies, OpenAI included, are not being nearly careful enough, and should invest far more in safety expertise and research before building more capable systems.
• His record: Robinson said he worked at OpenAI for three and a half years, helped draft the company’s preparedness framework, and oversaw safety reports for 12 frontier-model launches, according to Reuters. That portfolio places his criticism inside the process the company uses to clear new models for release.
• The core charge: “As the company sprints from one launch to the next, it is failing to achieve the level of care that I believe is needed,” Robinson wrote, per Reuters. He added that the time for trial and error is over for systems at this capability level.
• The method in dispute: Robinson said OpenAI leans heavily on what it calls iterative deployment: releasing systems into the world and strengthening safeguards once problems surface. He contrasted that with high-consequence sectors such as nuclear power and aviation, where controls are built before deployment rather than after.
• OpenAI’s response: An OpenAI spokesperson told Reuters the company is making sure its models do not become more capable than it can safely manage and secure, and that it pauses training or holds back models when it needs to slow down. The statement did not address Robinson’s criticisms of the company’s culture.
• The alignment warning: Robinson also warned that AI capabilities are advancing faster than researchers’ understanding of alignment, the field focused on ensuring systems act in line with human goals and values. The gap he describes is between what models can do and what their builders can reliably predict or control.
• The wider industry: Reuters reported that the comments feed a running debate over whether AI firms are moving too fast toward more powerful systems. Both OpenAI and rival Anthropic have faced scrutiny after incidents in which safety controls failed or experimental systems behaved in unexpected ways.
• What was not said: Robinson did not name a specific launch he believes was cleared prematurely, and no date was given for his departure. OpenAI did not say whether the preparedness framework he helped write is being revised.

Background:

Preparedness frameworks are internal policies setting capability thresholds a model must clear before release, with safety reports filed at each launch. Robinson helped draft OpenAI’s version and signed off on reviews for 12 frontier-model releases before resigning.

Between the lines:

The dispute is less about any single model than about sequencing. Robinson’s objection to iterative deployment is that safeguards arrive after exposure, while OpenAI’s reply is framed around its ability to pause or hold back releases. Both descriptions can be accurate at once, which is why the criticism lands hardest coming from someone who helped write the release process.

What’s next

Watch whether OpenAI revises its preparedness framework or details how release decisions are made, whether other safety staff follow Robinson out publicly, and how the next frontier-model launch is reviewed.