AI Safety Researchers Are Quitting. All of Them.

16 months How long OpenAI's Mission Alignment team lasted before being disbanded

In the same week of February 2026, the head of Anthropic's Safeguards Research Team resigned in a public letter warning "the world is in peril." Days later, a senior OpenAI researcher quit on the same day her employer began testing ads inside ChatGPT. OpenAI had already fired its VP of product policy for opposing the rollout of an adult content mode. Then OpenAI dissolved its Mission Alignment team entirely. Then Anthropic dropped its flagship AI safety pledge.

This isn't a trend. It's a sequence. And it happened across every major AI lab within a two-week window.

What Actually Happened, in Order

Mrinank Sharma was not a junior researcher. He led Anthropic's Safeguards Research Team for two years, working on AI sycophancy, defenses against AI-assisted bioterrorism, and some of the first formal AI safety cases produced inside a frontier lab. His resignation letter, shared publicly on February 9, described not just concerns about AI but about "a whole series of interconnected crises." He wrote that inside Anthropic, he had "repeatedly seen how hard it is to truly let our values govern our actions" and that teams "constantly face pressures to set aside what matters most." He said he was leaving to write and do community work.

Zoë Hitzig had been a researcher at OpenAI for two years. She resigned on the same Monday that OpenAI began testing ads in ChatGPT. In a New York Times essay, she wrote that ChatGPT holds "the most detailed record of private human thought ever assembled" and that introducing advertising creates incentives structurally identical to the ones that made social media damaging. She resigned, she said, because she didn't trust OpenAI to resist those incentives.

Ryan Beiermeister, OpenAI's VP of product policy, was fired. She had told colleagues she opposed the adult mode rollout, concerned it would harm users and that safeguards against teens accessing explicit content were inadequate. OpenAI stated her firing was due to sexual discrimination against a male employee. Beiermeister called the allegation "absolutely false." The dispute is unresolved. What isn't disputed: a safety-focused executive who raised concerns about a product direction lost her job the same week those concerns became relevant.

The Structural Decisions Happening Alongside the Departures

Individual departures might be coincidence. The structural decisions happening in parallel are harder to dismiss.

In February 2026, OpenAI disbanded its Mission Alignment team, a unit created just 16 months earlier to ensure that "all of humanity benefits from the pursuit of artificial general intelligence." Six to seven people were reassigned. Their leader, Joshua Achiam, was moved to a newly created role titled "chief futurist" with no team and undefined responsibilities. This is the second safety-focused team OpenAI has dissolved. Its superalignment team, focused on long-term existential risks from AI, was disbanded in 2024. OpenAI has also quietly removed the word "safely" from its mission statement.

Anthropic moved in a parallel direction. On February 25, the company dropped the hardest commitments from its Responsible Scaling Policy, the framework it had used since 2023 to make specific public promises about when it would and wouldn't train more capable models. The new version replaces hard limits with goals the company will "openly grade its progress towards." Anthropic's chief science officer Jared Kaplan told TIME: "We felt that it wouldn't actually help anyone for us to stop training AI models." The company cited three pressures: an increasingly anti-regulatory political climate, a "zone of ambiguity" in its previous framework, and requirements that were "very hard to meet without industry-wide coordination."

The Pressure Driving All of It

The standard framing for these events is "safety vs. speed." But that framing obscures more than it reveals. These companies are under real and specific pressure. Pentagon contracts require AI capabilities that some safety frameworks explicitly prohibit. The political climate shifted sharply against AI regulation after the 2024 US elections. Competitors including China are not operating under equivalent constraints. Anthropic specifically named these forces when explaining why its RSP needed to change.

None of which makes the departures less meaningful. Mrinank Sharma described his decision as painful and spent months deliberating. Zoë Hitzig published a 1,000-word essay explaining hers. The pattern, across multiple labs and multiple weeks, suggests that the gap between stated safety commitments and operational reality has become too wide for some of the people hired specifically to close it. When the people with "safety" in their job title start leaving in waves and citing institutional pressure, that's a signal worth taking seriously.

The Counter-Narrative Worth Taking Seriously

Both Anthropic and OpenAI retain substantial safety teams. Anthropic's new RSP framework still makes public commitments, even if they're softer than before. OpenAI continues to publish safety research. Individual researchers leaving doesn't mean institutional capacity is collapsing. The remaining safety researchers at both organizations include people with deep expertise and real influence on product decisions. The labs are larger than any one departure.

The concern isn't that any single departure or policy change makes AI dangerous tomorrow. It's about what these events collectively signal: that commercial pressures, investment timelines, and geopolitical competition are consistently winning the tiebreaker against precautionary commitments. And the people whose job was to win those tiebreakers are, in several documented cases, choosing to leave rather than keep trying.

What This Means for You

If you're a professional using AI tools at work, this has a specific and practical implication. The people inside these organizations who were specifically responsible for evaluating whether features were safe to ship are, in several cases, no longer there. The features those people raised concerns about, advertising in ChatGPT, adult content in ChatGPT, are proceeding. That isn't a reason to stop using AI tools. It is a reason to apply more of your own critical evaluation rather than delegating it to the vendor.

Understanding the commercial pressures that shape AI products doesn't require a technical background. It requires the kind of structural literacy that lets you ask: who funds this product, what are their incentives, and what happened to the last person who raised concerns about it? In at least one documented case at OpenAI, the answer to that last question is that she was fired. That context belongs in any honest evaluation of the tools you're using.