The person who wrote OpenAI’s safety reports has resigned, and he says the AI industry’s culture is too broken for a few new rules to fix. David Robinson wrote the safety reports that came out with every major OpenAI model release. He resigned this week, according to The Verge AI, and he’s now making his case in an editorial in The Atlantic.
His message is blunt. AI labs are moving too fast and with too much confidence. They also don’t have enough layers of protection against their own mistakes.
What Robinson Is Saying
Robinson’s main argument isn’t about one model or one policy. It’s about mindset. He says Silicon Valley has run on “extreme confidence” and “perpetual sprints.” Labs keep building bigger models with “unimpeded optimism” that ignores or underestimates what could go wrong.
His fix is humility. He wants labs to look for answers outside the closed, move-fast-and-break-things world of tech. Specifically, he wants frontier labs to copy industries where a failure can be a disaster:
“Given today’s risks, frontier labs need to run like nuclear power plants or busy airports, with layers of redundancy and careful, time-consuming planning, so that the occasional and inevitable human error does not open a door to disaster.”
That comparison stands out. Nuclear plants and airports don’t assume people won’t make mistakes. They assume people will. Then they build systems so that no single slip turns into a catastrophe.
Why His Exit Carries Extra Weight
Plenty of outsiders have criticized AI labs. Robinson’s warning matters more because of the job he had. Safety reports, often called system cards, tell the public what a model can do, what risks testers found, and what the company did about them. The person who writes them can see the difference between what got tested and what actually shipped.
The Verge AI admits that some skepticism is fair. Many of the people sounding alarms now helped build the technology they’re warning about. But that inside knowledge is the reason to take their warnings seriously, not a reason to dismiss them.
A Growing Exodus
Robinson isn’t alone. He’s the latest in a string of researchers and safety staff who’ve left top AI labs and spoken out:
- Jacob Coxon seems to have started the wave. He quit Anthropic and then said publicly that AI “could kill us all by the end of the decade.”
- Robert O’Callahan, Bilal Chughtai, and Josh Engels left Google DeepMind.
- Joe Benton left Anthropic.
That’s people leaving three of the biggest frontier labs. When safety staff at rival companies reach similar conclusions on their own, it’s hard to write it off as one person’s grudge.
What It Means If You Build on AI
If the people who test frontier models think the process is rushed, the models you’re putting into your products may carry risks nobody has fully mapped. A safety report is only as good as the time and culture behind it.
What should teams do now?
- Don’t treat a system card as a guarantee. Use it as a starting point, then run your own tests on your specific use case.
- Build your own backups. Robinson’s nuclear comparison works at a small scale too. Add human review, output checks, and fallbacks to high-stakes workflows.
- Watch who’s leaving. When safety staff leave a lab, its approach to releases often changes soon after. Pay attention to who’s quitting and what they say.
- Expect pressure from regulators. Public warnings from insiders tend to give lawmakers ammunition, so compliance rules could get stricter.
What Comes Next
The big question is whether labs will make real structural changes or just keep sprinting. Robinson argues that rules alone won’t fix this. The culture itself has to slow down. With OpenAI, Anthropic, and Google DeepMind competing as hard as ever, that’s a tough ask.
Still, every resignation adds pressure. If more safety staff keep quitting and going public, moving fast could start to hurt these labs instead of helping them. You can find more details in the original report from The Verge AI.