The Culture They Called Broken On The Way Out

Another OpenAI safety employee has resigned, and the resignation came with a claim attached: the company's "culture is broken." That phrase is not an isolated complaint. It is the latest entry in what reporting now describes plainly as a pattern, researchers leaving with public warnings rather than quiet departures. The pattern is the story now, more than any single person's grievance, because a pattern has a shape, and the shape tells you something the company itself is not saying out loud.

A genre, not an accident

One resignation with a public complaint is a personnel matter. A second is a coincidence. A pattern, which is what this now is, is something closer to an institutional signal, and institutions rarely send signals by accident. Safety researchers at OpenAI have developed, whether deliberately or not, a recognisable exit format: leave, then say on the record that the culture inside was not what it claimed to be from outside.

That format has value precisely because it repeats. A single whistleblower can be dismissed as disgruntled. A sequence of departures, each independently choosing to go public rather than sign off quietly, is much harder to wave away, and the people doing the leaving know that. The warning works because it is not the first warning, and it will not be mistaken for a one-off grievance by anyone paying attention to the run of headlines.

There is also a quieter effect at work. Each new departure makes the next one easier to credit, because the audience has already been primed by the last one. A journalist, a regulator, or a rival researcher reading the latest resignation does not start from zero. They start from a running total, and the total is doing persuasive work that no individual statement could do alone.

The claim underneath the claim

"Culture is broken" is a vague phrase on its own. What gives it weight is who is saying it and when. These are not disgruntled sales staff or departing marketing hires, they are the people whose job was specifically to sit inside the organisation and watch for exactly the kind of drift they are now describing from outside it. When the safety function itself produces a string of people saying the internal culture does not match the external promise, the complaint is not generic, it is a report from the one department built to notice.

That is the real content of these resignations. Not personality clashes, not pay, but a claim that the mechanism meant to catch problems has itself become a problem, and that the people inside it concluded the only effective channel left was the public one, delivered on the way out rather than through whatever internal process exists for raising concerns. If the internal process worked, there would be less reason to repeat the same warning in public, resignation after resignation.

It is worth sitting with the asymmetry here. A safety team exists to say uncomfortable things to the people above it. If the team's own members feel the only way to say an uncomfortable thing and have it heard is to say it on their way out the door, then the internal channel has already failed before the public statement is ever written. The resignation is not the failure, it is the record of a failure that happened earlier and privately.

A company quietly changing its own story

These departures land alongside another piece of reporting worth sitting with: OpenAI, apparently, isn't trying to build "magic intelligence in the sky" any more. That phrase, grand and slightly embarrassing in retrospect, used to sit close to the company's own description of its ambition. Reporting now suggests the company has moved away from it, quietly, without the kind of announcement that accompanied the original framing.

Put the two threads together and a shape appears. Safety staff who were hired, in part, to manage the risks of exactly that kind of grand ambition are leaving and saying the culture around them is broken, at roughly the same moment the company is said to be stepping back from the ambition itself. Either the retreat from "magic intelligence in the sky" is a genuine recalibration that the departing researchers haven't been told about or don't trust, or the public-facing story and the internal one have simply stopped being the same story, and the resignations are the visible seam where they no longer line up.

A company does not usually drop its founding rhetoric and lose a run of safety staff in the same stretch of time by coincidence. The two are more plausibly read as two symptoms of the same underlying condition: an organisation whose ambitions and whose internal management of those ambitions have come apart, with the public only getting to see the exit wounds.

Compare the neighbours

It is worth noticing how differently other companies in the same industry are handling pressure and criticism right now. Amazon, facing data centre backlash, responded by saying it no longer uses NDAs, a direct and specific answer to a specific complaint. Whatever one thinks of the underlying issue, that is a company meeting criticism with a concrete, checkable claim about its own practice, one that can be tested against future behaviour.

OpenAI's pattern looks nothing like that. There is no concrete counter-claim being offered against "culture is broken," no specific policy named as the thing that changed or the thing that will change. The warnings accumulate, and the company's public response to the accumulation has mostly been silence or generality, which is itself informative. A company that believes a departing researcher is wrong usually says, specifically, what they got wrong. Repetition without rebuttal starts to look less like restraint and more like an absence of a good answer.

Why the warning is the product now

None of this requires believing every departing researcher is right about everything. It requires noticing that the warnings have become the most legible output of the safety function, more legible than any internal report, because internal reports do not get read outside the building and public resignations do. If the people inside the safety team have concluded that the only reliable way to make a concern travel is to attach it to their own departure, then the departure itself has become the delivery mechanism, and the warning is the actual work product, not a regrettable side effect of someone leaving.

That is not a healthy sign for the organisation being warned about, whatever one makes of the underlying technical disputes. A safety function that can only be heard on its way out has already stopped functioning as a safety function while its members were still employed. The resignations are not noise around the real story. At this point, for OpenAI, they are close to being the real story, and the quiet retirement of "magic intelligence in the sky" reads less like humility and more like an admission arriving by a side door, dressed up as a change of strategy rather than a change forced by the people who kept saying, on the way out, that something inside had broken.