OpenAI fired three safety researchers for what it claims was “a significant breach of trust” involving the handling of sensitive information. The researchers say that’s not what happened. And honestly? It doesn’t matter who’s technically right about the policy violation, because the message OpenAI just sent is crystal clear: if you work on AI safety there, keep your head down and your mouth shut.

Let me be direct about this. I don’t know exactly what Jasmine Wang, Tomek Korbak, and Mikita Balesni did or didn’t do with company information. Maybe they actually violated a policy. Maybe OpenAI’s investigation was thorough and fair. But the timing here is doing all the talking, and what it’s saying is ugly.

These weren’t random engineers. They were safety researchers, the people literally hired to worry about whether OpenAI’s technology might cause harm. And they got walked out right when questions about OpenAI’s safety culture are at their loudest. The company insists this wasn’t retaliation for speaking out about safety concerns, but come on. That’s exactly how it looks, and OpenAI is smart enough to know that’s exactly how it looks.

The Chilling Effect Is the Point

Here’s what matters more than the specifics of this case: every other safety researcher at OpenAI (and frankly, at every other AI lab) just watched three of their colleagues get fired. They read OpenAI’s public statement about “significant breach of trust.” They’re now doing the math on what happens if they raise uncomfortable questions too loudly or share concerns with the wrong people.

That’s what a chilling effect looks like. It doesn’t require explicit threats. It just requires visible consequences.

OpenAI has spent years telling us that safety is paramount. Sam Altman has given countless interviews about the enormous responsibility the company feels. They created a Superalignment team (which dissolved after its leaders left). They published papers about their commitment to careful deployment. And now, when three safety researchers allegedly step out of line, the company’s response is to fire them and issue a terse statement about policy violations.

You can’t have it both ways. You can’t position yourself as the responsible AI company that takes safety seriously while simultaneously creating an environment where safety researchers are afraid to do their jobs properly.

What Safety Work Actually Requires

Good safety research requires uncomfortable conversations. It requires people who will push back on product timelines, question executive decisions, and sometimes share concerns with people who can actually do something about them. If your safety team is worried that flagging a real problem might get them fired for “mishandling sensitive information,” congratulations, you don’t actually have a safety team. You have a compliance department.

The researchers’ open letter warns about exactly this, and they’re right to be worried. When safety researchers start optimizing for not getting fired instead of actually identifying risks, the whole enterprise becomes theater. You get people who write careful memos that go into carefully managed channels where they can be carefully ignored.

OpenAI’s response to all this has been to double down. They posted on X (because of course they did) reiterating that the firings were totally justified and definitely not about safety concerns. They want us to trust that their internal processes are sound, that the investigation was fair, that everything was handled appropriately.

But trust is earned through transparency, not through press releases that amount to “trust us, we investigated ourselves and we were right.” If OpenAI actually wants to demonstrate that this wasn’t about suppressing safety concerns, they need to show their work. What was the actual violation? What was the process? Who made the decision?

The Pattern That Keeps Repeating

This isn’t happening in a vacuum. Jan Leike and Ilya Sutskever, who led OpenAI’s Superalignment team, both left the company earlier this year. Leike’s resignation thread on X was pointed: “Building smarter-than-human machines is an inherently dangerous endeavor. OpenAI is shouldering an enormous responsibility on behalf of all of humanity. But over the past years, safety culture and processes have taken a backseat to shiny products.”

That was months ago. OpenAI’s response to losing its top safety leaders was to restructure, make some organizational changes, and keep shipping products. Now three more safety researchers are gone, fired under circumstances that look an awful lot like retaliation, whether or not that was the intent.

At some point, the pattern tells you everything you need to know. When the people hired to slow you down keep leaving or getting pushed out, maybe the problem isn’t that you keep hiring the wrong people. Maybe the problem is that you never actually wanted to be slowed down.

What This Means for AI Safety Broadly

The AI safety field is small. Everyone knows everyone. When OpenAI fires three researchers under these circumstances, it sends ripples through the entire community. Other companies are watching. Other researchers are watching. And they’re updating their priors about what happens when you take safety concerns seriously enough to potentially upset your employer.

This is how you end up with an industry where the safety research that gets done is the safety research that doesn’t threaten product timelines or revenue projections. Where the risks that get studied are the ones that are safe to study. Where “alignment” becomes a buzzword that means “make sure the chatbot is polite” instead of “make sure this technology doesn’t cause catastrophic harm.”

I want to be wrong about this. I want OpenAI’s explanation to be the whole story. I want these firings to actually be about a clear policy violation that had nothing to do with safety advocacy. But I’ve watched this industry long enough to recognize the smell of safety theater, and this absolutely reeks of it.

If OpenAI wants to prove otherwise, they know what to do. Show the receipts. Explain the actual violation in detail. Demonstrate that your safety researchers can raise hard questions without fear of retaliation. Until then, we’re left with what we can see: three safety researchers fired, a company doubling down on its decision, and a whole field of AI safety workers taking notes about what happens when you speak up.

That’s not a safety culture. That’s a warning.