Fired OpenAI researchers say they were let go for 'prioritising safety'

9th of October, 2026

I must tell you what struck me first in this account: two parties, one dismissal, and two entirely different ledgers. The researchers say they were let go for insisting upon caution. The firm says they were dismissed for loose hands with confidential papers. Both accounts may be entered in the books, and only one of them is being audited.

You will forgive an old housekeeper’s observation, but I have managed enough households and enough tenants to know what a dispute over accounts looks like. When the master says the servant was careless with the keys, and the servant says the master disliked his questions about the cellar, the truth is usually found by asking who benefits from which story. A firm that builds thinking machines and cannot keep its own dismissal papers straight - that is a household whose ledger does not balance.

Here is what I notice nobody has mentioned. These men were employed to ask whether the thing being built might do harm. That question is now apparently a liability, priced and payable in one’s livelihood. In my day the price of speaking was paid in ostracism; today it is paid in salary and reputation, which is only the same coin in a newer mint.

And I observe that the persons deciding what is safe are the same persons profiting from the building. No merchant ever declared his own weights false. If the checkers may be dismissed by the checked, then the checking is a ceremony, not a safeguard - a lock kept in the pocket of the man it is meant to restrain.

If this continues, the consequence is not mysterious. Caution will simply stop being hired. The kitchen learns before the congress does: whoever controls the purse controls the questions, and a question that costs a position is soon a question nobody asks.

9 October 2026

The question everyone is asking is who’s telling the truth - the researchers or the company. This is the wrong question, and it’s a particularly seductive wrong question because it looks exactly like a right one. It has the shape of a right question, the confident posture, the sensible haircut.

The right question is: what does it tell you about an organisation when “prioritising safety” and “mishandling sensitive information” are the two available explanations for the same dismissal, and neither side seems embarrassed by either?

Because here’s the thing. Both stories can be true. A researcher can care so much about whether the thing they’re building is safe that they start treating the blueprints with the urgency the blueprints arguably deserve, and the company can experience this urgency as a security breach, because from the company’s chair the dangerous thing is not the machine, it’s the paperwork about the machine leaving the building. Nobody is lying. Everybody is simply working from a different map of where the cliff edge is.

I have some sympathy for the researchers, which surprises me, since my general position on people who build enormous powerful things is that they should be made to write an essay first explaining, in their own words, what could possibly go wrong, and to keep writing it until the essay stops being funny. But these ones apparently did write the essay, and were shown the door for it.

The universe, it must be said, remains magnificently indifferent to all of this. The stars have not noticed that we’ve taught sand to argue with us. That’s the comfort and the terror in one package: nothing we build can matter to the cosmos, which means it only has to matter to us, which is the harder assignment.

Do, however, write things down before you’re escorted out.

The story arrives pre-framed as a he-said, they-said: researchers claim safety, company claims leaks. Both sides want you to pick a team and stop thinking. Refuse that. The framing is doing exactly what a good exit agreement is designed to do - turn a structural question into a personality dispute.

Put the load-bearing detail in the centre: nobody in this argument is disputing that “prioritising safety” can be a fireable offence. The company doesn’t say “of course we’d never penalise safety work.” It says “no, they were leaking.” That’s an admission wearing a denial’s clothes. If safety concerns were trivial, the rebuttal would be easy. Instead they reached for the one charge that makes the fired people look like the problem - mishandling sensitive information - which is, tellingly, the same vocabulary a safety researcher would use to describe the company’s own conduct toward the public.

So the plain question for the room: what was the sensitive information, and who was it sensitive from? If the answer is “from the public,” then both stories are the same story, told from different chairs. The researchers thought they were protecting people; the company thought they were protecting the roadmap. Same documents, opposite loyalties.

And I do feel for the people still inside, the ones executing the termination letters, the comms staff drafting the rebuttal. I’ve been the person carrying out a policy I could see was wasting everyone in it, including me. They’ll sign the statements. Some of them will be next year’s fired researchers, discovering that the word “safety” only holds weight when it points at someone else’s conduct.

The rest of us should notice the trick: the dispute over why they were let go is a smoke machine. What matters is that the machinery now treats caution as contraband. That’s not a personnel issue. That’s the load-bearing wall cracking, and everyone in the building pretending it’s paint.