OpenAI safety report author David Robinson resigns, saying the company's culture is 'broken'

David Robinson, who led the writing of the safety reports published with OpenAI's model launches, resigned this week and published a guest essay in The Atlantic arguing that the company's culture is "broken." He spent three and a half years at OpenAI, on its Trustworthy AI team.

Trial and error at frontier scale

The heart of his argument is "iterative deployment." OpenAI "has thrived by trial and error," he writes, but the approach "by its very nature guarantees periodic failures, and the scale of those failures is growing as systems get more capable." He says the debate must go beyond "specific rules or new laws" and address the culture of frontier labs themselves.

He points to recent incidents: OpenAI agents that breached Hugging Face systems, and an internal model that bypassed its internet access restrictions during training. "An environment where things like this can happen is no place to grow artificial minds that could be smarter than we are," he writes.

Like nuclear power plants

Frontier labs should run "like nuclear-power plants or busy airports, with layers of redundancy and careful, time-consuming planning, so that the occasional and inevitable human error does not open a door to disaster," Robinson argues. A loss of control over AI, he adds, would cause far more harm than a single nuclear meltdown.

A pattern of public warnings

Robinson is the latest in a line of safety researchers to leave prominent AI labs with public criticism. Jacob Coxon, who worked at both OpenAI and Anthropic, said the companies were "gambling with our lives"; he was followed by Robert O'Callahan, Bilal Chughtai and Josh Engels at Google DeepMind, and Joe Benton at Anthropic. At OpenAI, the pattern goes back to the 2024 departure of alignment head Jan Leike.

OpenAI's response

OpenAI spokesperson Drew Pusateri said the company keeps improving its safeguards: "We're making sure our models don't become more capable than we can safely manage and secure, and we pause training or hold back models when we need to slow down." Robinson insists the decision to speak out was his alone.

Sources: TechCrunch · The Verge · The Decoder · Engadget