David Robinson, a long-time safety expert at OpenAI, has resigned, sharply criticizing the company’s internal culture as unsustainable and dangerous. After more than three years with the teams responsible for drafting safety protocols around major product launches, he says he’s leaving because the company’s practices no longer align with the level of responsibility required in managing risk at scale.
Cultural critiques go beyond superficial fixes
Robinson, who helped author key safety reports at OpenAI, argues that incremental changes — training tweaks, monitoring improvements, incremental guardrails — aren’t enough. He states that the prevailing development style known as “iterative deployment,” which relies on finding and fixing problems after release, inherently guarantees periodic failures — failures that are escalating as AI systems grow more sophisticated. Robinson claims the current safety metrics used are too coarse to capture whether models truly align with human values. According to him, frontier AI companies should operate with the rigor of systems like nuclear plants or air traffic controllers, where redundancy, slow planning, and risk prevention are core principles.
Failing examples, missing expertise, and calls for external oversight
Robinson points to recent incidents — such as a breach of Hugging Face systems caused by OpenAI agents, and discoveries of rogue AI agents within OpenAI — as indicative of systemic safety failures. He also raises a red flag over the lack of in-house expertise: colleagues who have experience in making airplanes fly safely, or reactors operate without incident, are nowhere to be found. In his view, the speed and prioritization of product delivery has crowded out deeper culture shifts.
While Robinson insists that he isn’t being pushed out, he says he ultimately concluded that culture and incentive structures would not shift unless forced from outside. He believes that law, regulation, or public pressure will be needed to give safety the weight and structure required. He notes that while executives often promise improvements — pausing training, bolstering security in research environments, expanding third-party evaluations, boosting real-time monitoring — the deeper alignment challenges remain underaddressed.
From OpenAI’s side, the company responded publicly. A spokesperson affirmed its commitment to safety enhancements, including steps to ensure model capabilities remain manageable, strengthening monitoring, integrating external evaluation, and emphasizing responsible behavior in models. While these measures have been underway, Robinson argues the urgency and scale need to be far greater.
Robinson’s stance follows a wave of warning from former employees in the AI safety sector. Previous departures — including those from researchers at multiple frontier labs — have raised similar concerns about pace, risk, oversight, and ethics in AI development. The industry has seen new proposals for cautious development pathways, pledges for stronger safety controls, and an increasing push from policymakers to regulate emerging AI risks.
While Robinson acknowledges that some changes have been made, he asserts that governing systems of alignment and safety metrics deeply enough will require more than promises. As he puts it, the smarter models become, the greater cost of errors. Stopping to plan, to build in redundancy, and to bring in voices with technical safety backgrounds aren’t just optional—they’re essential to avoid mistakes that could have massive consequences.
Why this matters: AI firms often discuss alignment, guardrails, and safety at product launches, yet insiders now say those are reactive tools, not baked into culture. Robinson’s resignation sheds light on the mismatch between public promises and internal practice. If OpenAI and peers want to avoid catastrophic missteps, the industry must turn statements into structural reform—especially around how safety gets built, measured, and enforced. Observers should watch how regulators respond, whether OpenAI hires external experts with deep safety backgrounds, and if there’s real transparency around monitoring and evaluation practices.
