The takeaway: A Google DeepMind research scientist has published the most detailed inside account yet of a frontier lab signing a military AI deal — a 54-minute-read resignation essay naming who pushed back, who stayed quiet, and which promises gave way. We spent the morning checking it: the essay's factual spine — the classified Pentagon contract, the quiet 2025 deletion of Google's anti-weapons pledge, the Pentagon's extraordinary campaign against Anthropic — is corroborated by independent reporting. The inside-the-building story is his alone, and he says so himself. Read together with the summer we've just covered, it completes a picture: governments deciding what labs may sell, and labs deciding what they'll sell to governments — with the pledges losing both times.

Who he is, and what he says happened

Alex Turner — known in AI-safety circles as TurnTrout, author of well-cited alignment work — joined DeepMind's scalable-alignment team in 2023 and resigned in June, Business Insider reported. His essay, "Why I Left Google DeepMind" (published Wednesday, and near the top of Hacker News since), states the thesis early: this is "the story of something larger: how powerful people and institutions failed, one after another, to keep their AI ethics promises in the face of pressure."

The triggering event is public record. In late April, the Pentagon signed a deal giving it Gemini on classified networks — part of a push, as NBC News reported, to sign contracts with the major AI companies on language allowing "any lawful use" of their systems. Reporting on the Google deal quoted "should not"-style contract language on surveillance and autonomous weapons — alongside a clause under which Google gets no right to veto lawful government use. "Aspirational language with no legal restrictions" is Turner's verdict on those terms, and his conclusion: "Google still signed a deal handing over their AI without restrictions against killer robots or mass AI spying. Google's contract restrictions were even weaker than OpenAI's. At that point, I couldn't stay at Google in good conscience, so I left."

That OpenAI comparison is his analytic judgment, but it has a documented basis: OpenAI's own announcement of its February agreement with the renamed Department of War describes contractually codified restrictions — no mass domestic surveillance, human control over weapons where law requires it — which analysts criticised as loophole-prone but which exist in the contract. Google's reported equivalents are "should not" statements — the gap that is Turner's core complaint.

And the prehistory is settled record too: in February 2025, Google removed from its AI principles the 2018 commitments not to build AI for weapons or norm-violating surveillance — in a blog post co-authored by Demis Hassabis. Turner sets two quotes side by side — the deletion, and a later Hassabis interview insisting nothing changed about the principles — and writes: "Both cannot be true."

What he did, and what he says the famous names didn't

By his own account — and he flags the limits of it himself: "I never quote anyone's private words without permission… You can't verify my characterizations, so weight them accordingly" — Turner spent the spring escalating through every channel available to a research scientist. An internal petition to chief scientist Jeff Dean (its existence was reported by The New York Times). A 25-page "Red Line and Oversight Framework" drafted on vacation days and reviewed by military-law experts. Amicus signatures. A direct message to Hassabis — who, he writes, told him to get the Framework evaluated by "two senior people working in gdm policy. I sent it to them. They left the message on read."

The essay's most striking passages concern the field's elder statesmen, and precision matters here. Of Stuart Russell — for a decade the loudest academic voice against autonomous weapons — Turner writes that at a February conference Russell agreed onstage to push his safety association toward a public statement and a member poll: "The statement and poll both vanished." (Turner notes Russell hedged with "time permitting", and that it was the association's interim executive director who supplied "an evolving list of reasons" for the silence.) Yoshua Bengio's office, he says, declined to make a statement without explaining why. Jeff Dean gets a deliberately mixed portrait — Turner calls him "Google's most outspoken executive" and credits him as the only one of the powerful figures he approached to take visible action, co-signing an amicus brief backing Anthropic, before adding, hedge intact: "I also asked him to use his immense leverage to stop Google from making its own unethical deal with the military, and I don't think he did. He remains at Google despite his pledge."

The Anthropic prequel, corroborated

The essay leans on a confrontation we now know shaped the whole spring: when Anthropic refused unrestricted military use, the Pentagon designated it a supply-chain risk — a label usually reserved for firms tied to foreign adversaries — and ordered agencies and contractors to drop it. Anthropic sued, twice, and won a preliminary injunction against the ban while the designation fight continues on appeal. Some of Turner's sharper specifics — notably a threatened Defense Production Act invocation — go beyond the public reporting; the ultimatum itself made the press. And there is another side: the Pentagon's technology chief has publicly framed Anthropic as the unreasonable party, an account with its own vocal doubters. What's not in dispute: every other lab the Pentagon courted signed, and Anthropic's stand cost it its federal business for months. Turner draws the lesson without romance: "When profit and pressure met ethical commitment at Google DeepMind, pressure won and pledges lost. When profit and pressure met ethical commitment at Anthropic, ethics won."

Why one essay matters

Resignations over principle are not rare; documented ones are. What makes this account load-bearing is that its checkable claims check out, its uncheckable claims are labelled as such by the author, and it isn't isolated: 600-plus Google employees petitioned against the deal, DeepMind's UK staff have reportedly moved to unionise with the Pentagon contract among the stated grievances, and a second, still-employed DeepMind researcher has published his own essay reaching similar conclusions. The labs spent this summer telling regulators — on our pages too — that democratic governments and frontier labs keep each other honest. Turner's essay is one researcher's evidence for the opposite: that when the pressure arrived, the honesty ran one way.