David Robinson, the architect of OpenAI's Preparedness Framework, quit this week and published an essay in The Atlantic saying the company's culture is broken.
Robinson spent 3.5 years at OpenAI running safety transparency work, the job of writing system cards and publishing risk disclosures that are supposed to tell the public what the models can safely do. He signed off on safety reports for 12 frontier-model launches. That's not a minor role. The Preparedness Framework he helped build is the internal system that's supposed to catch catastrophic risks before a model ships. When a person in that position walks, and does it with a published verdict, it matters.
His essay's core complaint is cultural. Leading AI labs, Robinson argues, are places where shipping fast routinely wins out over getting it right. He invokes aviation and nuclear safety as contrasts, fields where reliability engineering carries life-or-death stakes and nobody gets applause for moving fast and breaking things. The line that's already being quoted: "the time for trial and error is over."
The timing makes the resignation heavier. Robinson's exit comes four days after OpenAI fired three safety researchers, Jasmine Wang, Tomek Korbak, and Mikita Balesni, for allegedly sharing confidential information with an outside AI safety organization. OpenAI framed it as a matter of handling sensitive material, not whistleblowing. Believe that distinction or not, three people left the safety team the same week its transparency lead walked out the door.
This isn't Robinson's first doubt in public. In September, he posted on X that he agreed people at OpenAI were "starting to" grasp what highly capable AI models imply, but he didn't know whether things were changing fast enough. He expressed some doubt.
The broader pattern matters here. OpenAI dissolved its Preparedness team months ago. It folded safety into the research organization under VP Mia Glaese in July. At least six senior safety figures have exited in roughly two years. The company is in the middle of preparing for a public listing, which means lawyers and bankers are poring over exactly the kind of risk disclosures Robinson used to author. One doesn't shed safety leadership and then ask the world to trust the risk documents it produces.
Robinson isn't just walking away either. He says he'll work from outside the company to strengthen safety incentives. That's a different kind of commitment, not a quiet exit, but a signal that the problem he's identifying is structural enough that it requires pressure from outside.
I notice because the usual interpretation of these departures is that safety work is thankless. It is. But Robinson's specific role suggests something sharper: OpenAI doesn't appear to value having the function he performed. A company doesn't dissolve four safety teams in two years and lose its safety transparency lead the same week it fires three safety researchers if that function is thriving. It dissolves safety teams when speed matters more than it does.
Top comments (0)