When the people hired to slow a company down keep leaving instead, the slowing-down function has effectively been removed from the product.
That’s my read on the latest departure out of OpenAI. A former safety employee resigned and went public with a criticism that’s narrower and more useful than it first sounds: the company’s fast-paced culture, built around rapid development, raises the odds of failures. TechCrunch reported it. NEWSMAX picked it up. Beyond that, the specifics are thin — neither report settles what triggered the exit or what happens next.
I review AI tools for a living, which means I spend most of my time on things I can actually measure. Latency. Hallucination rates. Whether the agent you paid for can complete a five-step task without wandering off. Culture isn’t testable. I can’t run a benchmark on it. So when a safety resignation lands, my instinct is to treat it as unverified until something downstream confirms it.
But that instinct has a flaw, and this story exposes it.
You can’t audit what you can’t see
Nobody outside these labs gets to inspect the safety process. We don’t see the internal evaluations, the red-team findings, the arguments about whether a model is ready. We see a launch post and a system card. That’s it. Which means the only unfiltered signal available to the rest of us is what happens when someone on the inside decides the gap between what they were told and what they saw got too wide to sit with.
A resignation isn’t evidence. It’s a single person’s judgment, delivered without the context that would let you weigh it properly. Bad managers produce resignations. So do genuine structural problems. From the outside, the two look identical.
The specific complaint here is still worth sitting with, though. Speed increasing the chance of failure isn’t a moral claim about anybody’s intentions. It’s just how engineering works. Compress a release cycle and you catch fewer problems before shipping. That’s true for a CRUD app and it’s true for a frontier model. The difference is the blast radius.
The pattern problem
What makes this one harder to dismiss is that it isn’t isolated. Jacob Coxon, a researcher who has worked at both Anthropic and OpenAI, resigned and said on X that the two companies are “gambling with our lives.” He put the chance of AI killing all humans above 10 percent. He also said he thinks Anthropic is doing its best, and that there’s still no plan for solving alignment in superintelligent systems. A separate Anthropic researcher has warned that AI could kill everyone by the end of the decade.
I’ll be straight about where I land on those numbers. I don’t know how anyone calibrates a 10 percent extinction estimate, and I’m skeptical of figures that can’t be checked against anything. The end-of-decade timelines read to me like conviction, not forecasting.
Set the numbers aside and something more mundane is left. Multiple people with direct access to the development process have chosen to leave and say so publicly. Separately. From different organizations. People generally don’t torch a frontier-lab résumé line for attention. That’s a costly signal, and costly signals are worth more than cheap ones.
What this actually means for anyone building on these tools
Here’s where I think the practical reading sits for developers and teams shipping on top of these APIs. Not panic, not dismissal.
- Treat internal safety as unverified. If your product depends on a model behaving responsibly in edge cases, build your own guardrails. Don’t assume the provider’s evaluation covers your use case, because you can’t see what it covered.
- Watch release cadence as a risk signal. Faster shipping is good for your roadmap and bad for your confidence in what’s been tested. Both of those are true simultaneously.
- Keep your own evaluations. Model updates ship quietly. If you’re not testing your critical paths against each version, you’re trusting a process you have zero visibility into.
- Don’t outsource judgment to vibes. A company’s safety messaging is marketing. Its safety record is behavior over time. Track the second one.
My verdict
One resignation proves nothing. A run of them, across labs, from people in safety roles, pointing at the same tension between speed and caution — that’s a pattern worth logging, even without the details to confirm it.
What bothers me most isn’t the claim that any particular culture is broken. It’s that there’s no mechanism for the rest of us to find out whether it is. No external audit, no published evaluation standard, no independent body with access. We’re left reading exit statements and inferring backward, which is a terrible way to assess systems this consequential.
Until that changes, resignations are the closest thing to an audit log we get. That’s not a reason to trust them completely. It’s a reason to be annoyed that they’re all we have.
🕒 Published: