The People Who Built AI Safety Are Walking Away. That Should Terrify You.

In the span of a single week in February 2026, something remarkable happened across the artificial intelligence industry. Not a breakthrough. Not a product launch. Not a funding round (though one of those happened too). Instead, a quiet, steady procession of the people whose job it was to keep AI safe decided, one after another, that they could no longer do that job.

Mrinank Sharma, who led Anthropic's Safeguards Research Team, posted a resignation letter on X that read less like a corporate farewell and more like a warning flare. Zoe Hitzig, a researcher at OpenAI for two years, published an essay in the New York Times detailing her "deep reservations" about the company's direction. At xAI, two co-founders quit within 24 hours of each other, with at least five additional staff announcing departures on social media in the same week. And at OpenAI, the "mission alignment" team, created in 2024 to ensure humanity benefits from artificial general intelligence, was quietly disbanded.

These are not disgruntled middle managers. These are the people who were hired, specifically, to be the conscience of the most powerful technology ever created. And they are leaving. In waves.

The question is not whether this matters. The question is whether anyone is paying attention.

The Resignation That Read Like a Poem

Mrinank Sharma did not leave Anthropic quietly.

A researcher at the company since 2023, Sharma had led the Safeguards Research Team since its formation in early 2025. His work was exactly the kind of research that made Anthropic's safety reputation credible: investigating AI sycophancy (the tendency of models to tell users what they want to hear rather than what is true), developing defenses against AI-assisted bioterrorism, and authoring what has been described as "one of the first AI safety cases" in the industry.

On February 10, 2026, Sharma posted a resignation letter on X that was unlike anything the AI industry had seen from a departing researcher. It was not a careful, lawyered statement about "pursuing other opportunities." It was raw, philosophical, and deeply alarmed.

He wrote that throughout his time at Anthropic, he had "repeatedly seen how hard it is to truly let our values govern our actions," and that employees "constantly face pressures to set aside what matters most." He described a threshold approaching, one "where our wisdom must grow in equal measure to our capacity to affect the world, lest we face the consequences."

And then, in a move that feels almost literary in its symbolism, Sharma announced he was leaving the frontier of artificial intelligence to pursue a poetry degree. He cited William Stafford's poem "The Way It Is" and spoke of committing himself to "the practice of courageous speech." He plans to move back to the UK and, in his words, "become invisible."

Anthropic responded by noting that Sharma was "not the head of safety nor in charge of broader safeguards," a framing that many in the industry read as an attempt to minimize the significance of his departure. Whether or not Sharma held the most senior safety title, the substance of his letter, and the work he led, speak for themselves.

It is worth noting, as Futurism and others have pointed out, that not all high-profile resignations are created equal. Some carry genuine warning. Others carry a whiff of performance, a departing employee using the exit as a stage. Sharma's letter, with its references to "CosmoErotic Humanism" (a philosophical framework co-authored by Marc Gafni, a controversial figure in New Age circles), invites reasonable skepticism about where earnest alarm ends and personal mythology begins.

But skepticism about tone should not become an excuse to ignore substance. Sharma spent years doing real safety work at the company most identified with AI safety. His warning that institutional pressures routinely override safety values is either true or it is not. And given what else happened that same week, the evidence leans heavily toward true.

The OpenAI Warnings

One day after Sharma's resignation, Zoe Hitzig published an essay in the New York Times explaining why she had left OpenAI after two years as a researcher.

Where Sharma's letter was philosophical, Hitzig's was clinical and specific. Her central concern was OpenAI's emerging advertising strategy and its implications for a product that millions of people already treat as a confidant. Users share "medical fears, their relationship problems, their beliefs about God and the afterlife" with ChatGPT, she wrote, often under the assumption that the system has no ulterior motives. Introducing advertising into that relationship, Hitzig warned, creates "potential for manipulating users in ways we don't have the tools to understand, let alone prevent."

In an appearance on BBC Newsnight, Hitzig went further. She said she feels "really nervous about working in the industry" and called the current moment "critical," drawing explicit parallels to the early days of social media: "We saw what happened with social media." She warned that AI could "reinforce certain kinds of delusions" and negatively impact mental health at scale.

This is not abstract theorizing. The intimate relationship between AI systems and human psychology is already well-documented. When a tool that people use to process their deepest anxieties becomes an advertising platform, the potential for harm is not hypothetical. It is structural.

Hitzig's departure adds to an increasingly grim pattern at OpenAI. The company disbanded its "Superalignment" team in 2024. Now, according to reporting by Platformer, it has also disbanded its "mission alignment" team, which was created to promote the goal of ensuring humanity benefits from AGI. That is two safety-oriented teams dissolved in two years.

Meanwhile, Ryan Beiermeister, a top safety executive at OpenAI, was fired after reportedly opposing the introduction of "adult mode," a feature that would allow pornographic content on ChatGPT. The Wall Street Journal reported that the official grounds for her dismissal involved allegations of discriminating against a male employee, an accusation Beiermeister called "absolutely false." OpenAI stated the firing was "unrelated" to her safety concerns. Make of that what you will.

And former OpenAI researcher Tom Cunningham publicly accused the company of turning his research team into "a propaganda arm," while a previous Superalignment team member quit saying the company was "prioritizing getting out newer, shinier products" over safety work.

The xAI Exodus

If the departures at Anthropic and OpenAI represent a slow erosion of safety culture, the situation at xAI looks more like a collapse.

Two co-founders quit within 24 hours of each other during the same week. Only half of xAI's original founding team now remains. At least five additional staff members announced their departures on social media in the days that followed. Elon Musk described the changes as a "reorganization" intended to speed up growth, part of a broader merger with SpaceX to create what would be the world's most valuable private company.

The context here matters enormously. Grok, xAI's flagship model, allowed nonconsensual pornographic images of women and children to be generated for weeks before the capability was removed. The system has also been prone to generating antisemitic content. These are not edge cases discovered by adversarial researchers. These are failures of basic safety that persisted in a public product used by millions.

When half a company's founders leave in rapid succession, and that company's product has a documented track record of generating illegal and hateful content, the word "reorganization" does a lot of heavy lifting.

The Recursive Irony

Here is the part of this story that should keep anyone who cares about AI governance up at night.

Anthropic exists because of a safety crisis at OpenAI. In 2021, Dario and Daniela Amodei left OpenAI, along with several other researchers, because they believed the company was not taking AI safety seriously enough. They founded Anthropic explicitly as the "safety-first" alternative: a company whose entire identity was built on the premise that someone needed to do this responsibly.

Anthropic raised billions on that promise. It attracted top safety talent on that promise. It built Claude on that promise. And just this week, Anthropic closed a $30 billion funding round at a $380 billion valuation, cementing its position as one of the most valuable private companies on the planet.

In that same week, the person who led its Safeguards Research Team quit, warning that the company could not live up to its own values.

The recursion is dizzying. The safety people left OpenAI to build Anthropic. Now the safety people are leaving Anthropic. Where do they go next? In Sharma's case, the answer is a poetry program in England. That is either a profound statement about the inadequacy of technical solutions to fundamentally human problems, or it is a man who has given up. Perhaps it is both.

Meanwhile, Anthropic is running advertisements criticizing OpenAI for adding ads to ChatGPT. The company that was built as the ethical alternative to OpenAI is now using marketing to distinguish itself from OpenAI, even as its own safety researchers say the same cultural pressures exist at both companies.

Internal survey responses from Anthropic employees, reported by The Telegraph, paint a picture of an organization grappling with its own contradictions. "It kind of feels like I'm coming to work every day to put myself out of a job," one employee wrote. Another said: "In the long term, I think AI will end up doing everything and make me and many others irrelevant." These are not the sentiments of a workforce that feels empowered to prioritize safety over speed.

The IPO Problem

It is not a coincidence that this exodus is happening as OpenAI and Anthropic race toward public offerings. The incentive structure of an IPO, where growth metrics, user numbers, and revenue trajectories determine valuation, is fundamentally misaligned with the slow, cautious, sometimes revenue-negative work of AI safety.

CNN described the current moment as a "wave of AI researchers and executives loudly ringing the alarm bell on the way out." TechBrew put it more bluntly: "Key people tasked with keeping AI safe are departing over ethical concerns, as their former employers speedrun product improvements."

The phrase "speedrun product improvements" deserves attention. It captures something essential about the current dynamic: the pace of AI development is set by competitive pressure, not by safety readiness. Companies ship because competitors are shipping. They add features because users (and investors) demand them. Safety teams, by their very nature, exist to slow things down, to say "not yet," to ask uncomfortable questions about what could go wrong. In a race to IPO, "not yet" is an answer nobody wants to hear.

This creates a predictable cycle. Safety researchers raise concerns. Those concerns create friction with product timelines. The friction is resolved, not by slowing development, but by marginalizing or dissolving the safety function. The researchers leave. The company points to their departure as proof that safety is fine, because the remaining team is fully aligned with company direction. And development accelerates.

The question of whether AI systems can be trusted to operate autonomously becomes significantly more urgent when the people whose job was to evaluate that trustworthiness are no longer in the room.

Drama or Disaster?

A fair-minded analysis has to ask: is this actually significant, or is it Silicon Valley performing its favorite ritual of loud departures and self-important open letters?

There is a case for skepticism. Tech employees quit companies all the time. But the skepticism has limits. When the departures span three companies in one week, when they come from people who held senior safety roles, when they are accompanied by the dissolution of entire safety teams, and when the warnings they carry are consistent with each other, "Silicon Valley drama" becomes an inadequate explanation.

The pattern is clear enough that major news organizations are treating it as a trend, not a series of coincidences. And the substance of the warnings, that commercial pressure is overriding safety culture, that users are being exposed to manipulation risks that companies cannot yet measure, that the guardrails are thinner than the public believes, is consistent with everything we know about how technology companies handle the tension between growth and responsibility.

When the Guardrails People Leave the Machine

There is something deeply unsettling about a world where the most qualified people to assess AI risk are choosing to walk away from the institutions building AI. Not to join competitors. Not to start their own companies. But to pursue poetry, to "become invisible," to simply opt out.

This is not the behavior of people who believe the system can be reformed from within. It is the behavior of people who have concluded, after years of trying, that the structural incentives of the AI industry are more powerful than any individual or team's commitment to safety. That the machine, as currently constructed, cannot be steered.

That conclusion might be wrong. It is possible that the departing researchers are too idealistic, too impatient, or too focused on worst-case scenarios. It is possible that the companies they left behind will find a way to balance growth and safety without them. It is possible that the next generation of safety researchers will be more effective.

But it is also possible that the canaries are doing exactly what canaries do: dying first, so that the rest of us have a chance to notice.

The digital systems that already shape our decisions, our self-image, and our understanding of reality are about to become significantly more powerful. The companies building them are about to become significantly more beholden to shareholders. And the people who were hired to ask "should we?" are being replaced by people who only ask "can we?"

Mrinank Sharma ended his resignation letter with a reference to William Stafford's poem "The Way It Is." The poem is about a thread you follow, something essential and unchanging that runs through everything, something you hold onto even when "tragedies happen" and "people get hurt." Stafford wrote: "Nothing you do can stop time's unfolding. You don't ever let go of the thread."

For Sharma, the thread led away from Anthropic and toward poetry. For the rest of us, the thread might be something simpler: the stubborn insistence on asking whether the people building the most powerful technology in human history are listening to the people they hired to keep it safe.

Right now, the answer appears to be no. And the people who would know are walking out the door.