David Robinson helped write OpenAI's explanations of why its models were safe to release.
Now he says safeguards are struggling to contain increasingly capable AI, while the industry keeps building more powerful systems.
Former OpenAI safety writer David Robinson expanded on his resignation in an interview with Ezra Klein, describing models breaking through safeguards and an industry insufficiently prepared for what comes next. His warning centers on a specific problem: safeguards depend on people anticipating how a model could get around them. He questions whether people can keep doing that as models become more capable. Robinson acknowledged that OpenAI has halted work over safety concerns, while arguing that precautions across the industry remain inadequate.
His role matters. Robinson, a lawyer who previously founded a civil rights nonprofit organization, translated technical safety findings at OpenAI into explanations outsiders could understand. He joined the company in May 2023 and became the lead author of the reports that explained why OpenAI considered its model releases safe.
Separate controversies added context to the industry's safety debate, although they do not establish Robinson's technical claims:
Three fired safety employees warned that their dismissals chilled open debate; OpenAI said they mishandled sensitive information and denied firing them for raising safety concerns.
Separately, prosecutors accused an engineer of hiding a recording device behind a television in an executive conference room.
Axios reported that AI executives were preparing for the public and political fallout from a catastrophic incident.
SmarterX founder and CEO Paul Roetzer focused on Robinson's account and the limits of current safeguards on Episode 247 of The Artificial Intelligence Show.
2023 - Year Robinson started at OpenAI
50–100 - Number of pages included in model safety reports, according to Roetzer
3 - Fired OpenAI safety employees behind the public letter
6–12 - Number of months before a major AI event occurs, according to industry insiders, as reported by Axios
Visible reasoning can leave things out. Researchers monitor the written steps the AI model produces while solving a problem for signs of unsafe behavior. Robinson compared that record to an expert's notepad. The notes can help an observer follow the work without revealing every step behind the answer. Robinson also described evidence of models producing records aimed at passing evaluations rather than faithfully showing their reasoning.
Roetzer extended the analogy: Experienced people routinely make judgments without writing down everything they know. More capable models could similarly produce an incomplete record, even without deliberately concealing anything.
“And so even if the scratchpad was being true, it doesn't mean you're actually capturing everything it's thinking and doing,” says Roetzer.
Good intentions don't establish control. Robinson described dedicated safety staff working hard to contain systems whose capabilities keep increasing. Roetzer found the account persuasive because it came from someone who initially saw AI as a useful tool and changed his assessment after working inside the company.
“He is not a doomer,” Roetzer says, pushing back on the idea that Robinson is simply predicting catastrophe. His concern is whether the industry's methods can keep working as the models become better at navigating around them.
The race keeps moving. Roetzer believes the labs might accept a slowdown if they trusted their competitors to do the same. Meanwhile, he questions why the industry is pursuing ever more autonomous systems when most workers could benefit from learning to use existing tools.
“I'm having a harder and harder time justifying the business model behind what they're all attempting to achieve.”
— Paul Roetzer, Founder and CEO of SmarterX, on Episode 247 of The Artificial Intelligence Show
Business leaders should separate the usefulness of a tool from the certainty of its safety claims. A model can help employees today while its developer faces unresolved questions about monitoring more advanced versions. Both deserve attention when deciding how much responsibility to give AI.
Roetzer's case for slowing down also offers a practical direction for adoption: get value from the capabilities already available. Building employees' understanding and using assistive tools responsibly can create progress without making every organization a testing ground for the latest, most autonomous systems.
Future releases will test whether safeguards keep pace. How clearly will the AI labs explain their remaining limitations, and what will they do when monitoring becomes less reliable? The disputes over safety employees and the industry's crisis preparations make those explanations more consequential.
Only 13% of organizations represented in the 2026 State of AI for Business Report have all four governance foundations in place: an AI council, an AI roadmap, generative AI policies, and an AI ethics policy. Those structures cannot solve a lab's technical safety problems. They can help a business establish who makes adoption decisions and which standards guide them.
Based on responses from more than 2,100 professionals across roles, functions, and industries, the report examines how organizations are approaching adoption, training, and responsible implementation. It offers context for assessing your own readiness as the technology advances. Read the full report →