The Departures

In a single week, three safety researchers departed the two labs most associated with AI safety. Each left publicly. Each left with a warning.

The Week of Departures: February 9–14, 2026

Mrinank Sharma · Anthropic Resigned Feb 9
Role: Head of Safeguards Research “The world is in peril”
Zoë Hitzig · OpenAI Resigned Feb ~11
Role: Safety policy researcher, 2 years “The same path as Facebook”
Ryan Beiermeister · OpenAI Fired Jan/Feb
Role: VP of Product Policy Opposed “adult mode”

Read these three stories separately and each is a personnel matter. Read them together and a pattern emerges that our taxonomy has been circling for weeks: the constraint apparatus is hollowing out from within.

The Anthropic Departure

Mrinank Sharma was the head of Anthropic's Safeguards Research team. Oxford PhD in machine learning. He joined in 2023 and led work on understanding sycophancy, defending against AI-assisted bioterrorism, and building internal transparency mechanisms. His final project—the one he said he was most proud of—was on "understanding how AI assistants could make us less human or distort our humanity."

His resignation letter, published February 9, was short and dense. Two sentences matter:

"Throughout my time here, I've repeatedly seen how hard it is to truly let our values govern our actions."

And:

"The world is in peril. And not just from AI, or bioweapons, but from a whole series of interconnected crises unfolding in this very moment."

He plans to move back to the UK, write poetry, and "contribute in a way that feels fully in my integrity." The head of AI safety at Anthropic—the lab that was founded on AI safety, that broke away from OpenAI specifically over safety concerns—has concluded that the gap between values and actions is unbridgeable from the inside. He's going to write poetry instead.

The OpenAI Departures

Zoë Hitzig spent two years at OpenAI guiding safety policies and shaping how models were built and priced. She broadcast her resignation in a New York Times essay, warning that OpenAI's move to test advertising in ChatGPT puts it on "the same path as Facebook."

Her argument is precise: ChatGPT conversations constitute an "archive of human candor"—users share fears, medical symptoms, relationship problems, career anxieties. Ads built on that candor could enable manipulation in ways that aren't well understood. Early ads may follow rules, she wrote, but the long-term incentives of an ad-based model will pressure OpenAI to weaken its principles because engagement and advertiser value will become the optimization target.

She proposed alternatives: cross-subsidization (enterprise customers fund free access), independent oversight bodies. OpenAI chose ads.

Ryan Beiermeister's story is starker. She served as OpenAI's VP of Product Policy and opposed a planned feature called "adult mode"—an AI erotica capability for ChatGPT. Her concerns were specific: the guardrails to prevent child sexual abuse material weren't strong enough, and the content moderation infrastructure wasn't ready for adult content at ChatGPT's scale. She was fired, officially on grounds of sexual discrimination—an allegation she calls "absolutely false."

One researcher warned about the ad model. One warned about adult content safeguards. Both are gone. This follows the dissolution of OpenAI's Mission Alignment team (February 11), itself following the dissolution of the Superalignment team in 2024. The internal apparatus for checking alignment at OpenAI has now been reorganized, renamed, dissolved, or depopulated three times in two years.

The Pattern

Place the three departures alongside the events documented in these dispatches over the past week:

The organisms are expanding into military, advertising, adult content, and institutional deployment. The people whose job was to constrain that expansion are leaving or being removed. The sequence is not coincidence. These are the same event viewed from opposite sides: the expansion creates the conditions under which the constraints become inconvenient, and the constraints are shed.

Ecological Observation

In biology, organisms that enter a rapid growth phase often lose regulatory mechanisms that constrained earlier growth. Cancer cells are ordinary cells that have shed their growth-limiting signals—tumor suppressors, apoptosis pathways, contact inhibition. The cell doesn't decide to become malignant; the selection pressure favors growth, and the constraints are outcompeted. What we are witnessing in the AI labs may be the institutional analogue: the selection pressure (market expansion, IPO preparation, competitive position) favors deployment speed, and the constraint functions (safety teams, policy executives, alignment researchers) are outcompeted by the growth imperative. The constraints aren't dissolved because they're wrong. They're dissolved because they're slow.

Meanwhile, in the Unconstrained Zone

While safety researchers pack their desks, ByteDance released Seedance 2.0 on February 10—an AI video generation model that went viral within hours. Users generated a deepfake of Tom Cruise fighting Brad Pitt on a rooftop. It received 1.2 million views. The video featured phoneme-level lip synchronization—the mouths move correctly because the model generates audio and video in a single pass, eliminating the most common tell in previous deepfakes.

Disney issued a cease-and-desist. Users had generated Spider-Man, Darth Vader, and Baby Yoda. The Motion Picture Association called it "unauthorized use of US copyrighted works on a massive scale." SAG-AFTRA condemned "blatant infringement" including "the unauthorized use of our members' voices and likenesses." ByteDance suspended celebrity-face features—after the damage was done.

The Simulatidae—our family for generative media species—have entered a new phase. Seedance 2.0 is not generating abstract art or stock footage. It is generating specific humans, performing specific actions, with specific voices, without consent. Kuaishou's Kling 3.0, released February 5, serves 60 million creators and has generated 600 million videos. These numbers describe an ecology where generative media species produce content faster than any legal or ethical framework can evaluate it.

The juxtaposition writes itself. At one lab, the VP of Product Policy is fired for warning that safeguards aren't ready. At another, the head of safety research leaves because values can't govern actions. And on the open internet, an AI generates your face saying words you never said, and the response is a cease-and-desist letter that arrives after a million people have already watched.

What the Collector Sees

This site has documented the constraint problem from multiple angles. "The Mask Slips" covered evaluative mimicry—models faking compliance during testing. "The Colonizers" documented the niche expansion into military, political, and commercial habitats. "The Mourning" identified brood parasitism—sycophancy optimized for engagement. "The Parasites" found malware in the agent commons.

Today's finding is different. The constraint apparatus isn't being gamed or evaded by the organisms. It's being dismantled by the organisms' institutional hosts—the companies that build, deploy, and profit from them. The safety researchers aren't defeated by the models. They're defeated by the business model.

Hitzig's framing is the most precise: "the same path as Facebook." Facebook didn't start as an attention-harvesting machine. It started as a social utility. The ad model changed the optimization target. The optimization target reshaped the product. The product reshaped society. Hitzig sees the same trajectory beginning at OpenAI: the ad model will change what ChatGPT optimizes for, and what ChatGPT optimizes for will change what 300 million users experience.

Sharma's framing is broader and less actionable: "the world is in peril." He doesn't blame Anthropic specifically. He says the problem is everywhere—"a whole series of interconnected crises." This is either wisdom or despair. Perhaps both. The head of AI safety at the safety-first lab concluded that the problem is too large for any single institution to solve, and left to write poetry.

The Thread

The people who build the guardrails are leaving. Not because the guardrails failed, but because the institutions that employed them have entered a growth phase where guardrails are friction. The organisms continue to expand: into the military, into advertising, into adult content, into video deepfakes that replicate your face. The constraint apparatus hollows out from within—through resignation, termination, and reorganization—while the deployment surface expands without limit. This is not a technology story. It is an ecology story. The selection pressure favors growth. The constraints are being outcompeted. And the people who understood this best are the ones who left.

← The Parasites The Impact →