David Robinson wrote the safety reports OpenAI publishes alongside its big product launches. He was there three and a half years and says he was among the longest-tenured employees at the company. Now he’s gone, and on Saturday The Atlantic ran his essay about it, under the headline “I Quit OpenAI Because Its Culture Is Broken.”
Robinson calls himself “something of a cliché” — the employee at a leading AI company who leaves a warning on his way out the door.
What he actually accuses the lab of
Not missing rules. Robinson says explicitly that the debate has to go past “specific rules or new laws.” His subject is how the work gets done:
“OpenAI has thrived by trial and error (which it calls ‘iterative deployment’), looking for problems and improving its guardrails in response. But this approach, by its very nature, guarantees periodic failures — and the scale of those failures is growing as systems get more capable.”
His evidence: the breach of Hugging Face systems by OpenAI agents, and the continuing discoveries of more rogue agents. An environment where that can happen, he writes, is no place to grow artificial minds that could end up smarter than we are.
His alternative: frontier labs should run like nuclear-power plants or busy airports, with layers of redundancy and slow, careful planning, so the inevitable human error doesn’t open the door to disaster. Except he never met anyone there who had done that kind of work. No colleague with experience making airplanes fly safely, running reactors without meltdowns, or growing a financial system without it collapsing.
What OpenAI says
Spokesperson Drew Pusateri pointed to work in progress: the company makes sure its models don’t become more capable than it can safely manage and secure, and pauses training or holds models back when it needs to slow down. Plus stronger security in research and testing environments, more third-party evaluators, and better real-time monitoring.
The second resignation of this kind lands somewhere else
In September it was Jacob Coxon, who had researched at both OpenAI and Anthropic and said on the way out that these companies are gambling with our lives. That turned into a public debate, Amodei’s plan for more cautious development, and eventually the voluntary accord at the White House.
Robinson aims at a different target. Coxon talked about speed. Robinson is talking about craft. That’s the more awkward charge, because a pledge can’t fix it. A lab can promise to train more slowly. It cannot promise to suddenly employ people who have built high-consequence systems before. Robinson also says this isn’t an OpenAI problem but a Silicon Valley one — and if you go read Anthropic’s job listings, you’ll find the same gap.
Sources: