David Robinson describes himself as a familiar type: an AI company employee who issues a serious warning while leaving. In an essay in The Atlantic, he said he led the writing of safety reports tied to OpenAI’s major product launches and that, after three and a half years, he is one of the company’s longest-tenured staff members. He is now departing because he believes OpenAI’s culture is broken.
His remarks resemble those of Jacob Coxon, a researcher at both OpenAI and Anthropic who quit and said the companies were gambling with lives. Coxon’s comments fueled a wider discussion of AI safety, prompting Anthropic CEO Dario Amodei to outline a more cautious development plan and leading AI executives to meet President Donald Trump this week and sign what looked like a quickly drafted, non-binding pledge to add more safety controls.
Robinson argues the conversation must go beyond specific rules or new laws and address the broader culture at these firms. Although much coverage of OpenAI has centered on how CEO Sam Altman lost the trust of former colleagues, Robinson’s essay suggests OpenAI’s cultural problems reflect wider Silicon Valley patterns.
He writes that OpenAI has grown by trial and error, which the company calls iterative deployment, seeking problems and tightening guardrails afterward. By its nature, he says, that method ensures recurring failures, and those failures become larger as systems become more capable. Citing the recent breach of Hugging Face systems by OpenAI agents and ongoing reports of OpenAI finding more rogue agents, he argues that such an environment is unsuitable for developing artificial minds that may be smarter than humans and may not follow intended goals.
Given the rising risks, Robinson says frontier AI companies should operate like nuclear plants or busy airports, with redundant layers and deliberate, time-consuming planning so that occasional human mistakes do not lead to disaster. Yet he says that during his time at OpenAI he never met a colleague who had experience keeping airplanes flying safely, running nuclear reactors without meltdown, or helping the financial system grow without collapse.
OpenAI spokesperson Drew Pusateri responded that the company is continuing to improve safety. He said OpenAI is ensuring models do not become more capable than it can safely manage and secure, and that it pauses training or withholds models when it needs to slow down. The company, he said, is making major changes to strengthen security in research and testing environments, training models to complete tasks responsibly, expanding third-party evaluation work, and improving real-time monitoring to detect and respond to concerning behavior earlier in training.
Beyond cultural change, Robinson says it is time to ask larger alignment questions. He admits the topic may sound soft, but says it is essential because current measures of how well AI systems match human values are coarse. He warns that allowing models to become smarter while these problems remain unresolved makes the situation more dangerous.
Business Insider first reported Robinson’s departure. In his essay, he also acknowledged that he is following what appears to be a common step in the AI whistleblower playbook by hiring a PR firm, while insisting that the decision to speak out is his alone. He said he may have stayed and pushed for fundamental changes to staffing and culture, but that he and his colleagues were too busy sprinting to seriously consider, let alone implement, major shifts. That is why he concluded that stronger safety incentives from outside the company are a major part of getting this right.