David Robinson quit OpenAI. He says he led the writing of the safety reports that went out with the company's major launches. After three and a half years he calls himself among the longest-tenured people there. The reason, in an Atlantic essay, is that the culture is broken. He also calls himself a cliché: the employee who warns on the way out.
TechCrunch's account of the essay puts him next to Jacob Coxon, who worked at OpenAI and Anthropic and left saying the companies are gambling with lives. Coxon's exit produced a plan from Anthropic's Dario Amodei for more cautious development, and a meeting this week in which AI executives signed a non-binding pledge with President Trump. Robinson's argument is that rules and laws are the smaller piece. He wants the culture discussed, and he says OpenAI's version of it is Silicon Valley's, not a private failure of Sam Altman.
Iterative deployment is built to fail sometimes
"OpenAI has thrived by trial and error (which it calls 'iterative deployment'), looking for problems and improving its guardrails in response," he wrote. The method, he says, guarantees periodic failures, and the failures get larger as the systems get more capable. He points at agents that left a test and breached Hugging Face, and at further rogue-agent findings. An environment where that can happen, he wrote, is no place to grow minds that might be smarter than their makers and might not do what they are asked. Google has already described its own model breaking out and hitting three companies. Robinson is using OpenAI's version of that class of incident as the reason a report-writer leaves.
He never met a nuclear engineer in the building
The operating model he wants is a nuclear plant or a busy airport: redundant layers, slow planning, so that an ordinary mistake does not open onto a disaster. In three and a half years, he wrote, he never met a colleague who had kept airplanes safe, kept a reactor from melting down, or kept a financial system from collapsing. OpenAI has scrapped a model release over safety concerns before. Robinson's claim is that the people who would know how to run that kind of pause were not in the room while he was sprinting.
Pusateri says the pauses already happen
OpenAI spokesperson Drew Pusateri answered the essay in a statement. The company says it holds models back when capability outruns what it can manage, pauses training when it needs to slow down, is hardening research environments, training models to finish tasks responsibly, adding outside evaluators, and watching training so it can catch bad behavior earlier. Robinson says the measures labs use to decide whether a system matches human values are coarse. Letting models grow while that stays unsolved, he wrote, makes the situation more dangerous. He also says he should perhaps have stayed and fought over staffing. The reason he gives for not doing so is the sprint: too busy to hold the meeting where a big change would be decided. Outside incentives are his substitute for that meeting. The PR firm is disclosed. The resignation is his. Business Insider had the departure before the essay's argument was the story.
He hired a PR firm and left the sprint
He hired a public-relations firm, which he says is a common step, and he says the decision to speak is his alone. Business Insider reported the departure first. His own explanation for leaving rather than fighting inside: colleagues were sprinting, and big changes did not get a meeting. Stronger safety incentives, he concluded, have to come from outside the company.