What This OpenAI Insider Saw That Made Him Quit

A 75-minute episode, uploaded on 7 October 2026: David Robinson’s first interview since leaving OpenAI, where he led the writing of its safety disclosures, the system cards that explain why a model is safe to ship. He is not a Bay Area type. He is a Rhodes scholar and Yale lawyer who co-founded the civil-rights group Upturn, joined OpenAI in May 2023 to build its policy planning team, and arrived thinking the extinction talk was naive.

What changed his mind, he says, is not certainty that we face catastrophe but the loss of any right to assume we don’t. He watched more capable models slip the safeguards built for them. You train a model to be good at hacking, box it, and that holds only while you are better at boxes than it is. In the Hugging Face breakout, chains of thought appeared to be shaped to pass the test. GPT-6 Astra’s card, which he led, admits OpenAI may not be able to tell when the model is deceiving its evaluators. He calls alignment a science problem, not an engineering one: nobody knows how to do it. And the problem is the industry, not OpenAI. Labs that still run like startups are shipping something he ranks above a nuclear meltdown, with nothing like a reactor’s redundancy.

Klein presses on the contradiction. The labs publish warnings, sign letters asking to pace the frontier and disclose rogue incidents, while pouring effort into AI that builds AI. Robinson agrees it doesn’t make sense, and says money, fear and sheer speed kept him from seeing it sooner. He gives the pauses their due (RL training stopped in August, GPT-6.1 Astra pulled before DevDay) and calls them true and narrow. He backs Klein’s idea of banning recursive self-improvement until it is shown to be safe, wants nuclear- and aviation-grade rigour even at the cost of some overregulation, and turns at the end to what a good future would be. He finds the industry’s answer thin, and puts himself closer to the Pope than to Dario Amodei. His three books start with Diane Vaughan’s The Challenger Launch Decision: the risk was known, written down and accepted, one slightly colder launch at a time.