This is a July 21, 2026 first-person essay in Transformer by Alex Turner, who worked as a research scientist at Google DeepMind until June 9, 2026. It is a participant account of an internal campaign against Google's classified Pentagon AI contract, and an argument that voluntary commitments and insider influence failed as governance mechanisms. It is a first-person opinion piece by an interested party and is treated here as testimony and argument, not as an independent record of Google's decision-making.
The commitments at issue
Turner sets out a chain of commitments he argues the Pentagon deal broke.
- 2014 acquisition. When DeepMind's co-founders sold to Google, the acquisition agreement specified — according to Google DeepMind CEO Demis Hassabis — that "no technology coming out of DeepMind will be used for military or intelligence purposes."
- 2018 pledge. The company publicly pledged to "neither participate in nor support the development, manufacture, trade, or use of lethal autonomous weapons." Google's original AI principles that year committed the company not to "design or deploy AI" in weapons "whose principal purpose or implementation is to cause or directly facilitate injury to people," nor for surveillance "violating internationally accepted norms."
- February 4, 2025 reversal. Hassabis co-authored a post announcing updated principles that removed the weapons and surveillance prohibitions. Turner's central factual charge is a contradiction: in an interview after the update, Hassabis said "nothing's changed about our principles," having announced the changes himself. Turner writes: "Both statements cannot be true." He states he raised the discrepancy in an internal channel and that many colleagues expressed disappointment.
See Autonomous Weapons.
The 2026 contracting round
Turner's account of the industry-wide sequence: the Pentagon pressed AI companies to hand over their models for "all lawful use," which he glosses as no restrictions against lethal autonomous weapons or domestic mass spying. Anthropic refused, and "the government retaliated by labeling the company 'a supply chain risk,' damaging its enterprise business" (see Anthropic v. United States (Pentagon ban challenge)). OpenAI signed, claiming its deal protected the same red lines Anthropic had insisted on, though "some analysts concluded that OpenAI's contract language contains wide loopholes." Google then signed a deal Turner describes as reportedly carrying "even fewer restrictions than OpenAI's contract."
His own intervention was a 25-page proposal containing contract language and oversight mechanisms, which he describes as a principled counteroffer Google could have stood by and says military- and surveillance-law experts praised. He sent it to Hassabis, who routed it to senior policy staff; after an initial review by relatively junior staff, "those senior staff never got back to me." He learned the deal had been signed "at 11:45 pm via a Signal group," adding that "Google never announced the deal internally."
The argument about individual leverage
A substantial part of the essay concerns Jeff Dean, Google's chief scientist, whom Turner identifies as the person with the most leverage: Google's 30th employee, developer of key algorithms, co-lead of the Gemini effort, and someone whose "departure would be a disaster for the company."
Turner records what Dean did do — tweeting that the DHS killing of Alex Pretti was "absolutely shameful," tweeting against AI for lethal autonomous weapons or mass surveillance of Americans, and co-signing an amicus brief backing Anthropic against the Pentagon, which Turner says he solicited and which "broke from the company line." He credits this: "I respect that." What he sought and did not get was a threat to resign: "I wanted him to threaten to walk if Google signed away its AI to the Pentagon — as I think his pledge demands. Dean evidently did not walk." His inference is stated as such: "I'm comfortable guessing that Dean did not even threaten to walk. If he had, I would expect the classified deal to contain at least some binding provisions."
Turner then generalizes into an argument about what a pledge obliges when it becomes costly. He allows the possibility that Dean and Hassabis did everything they could and failed, and argues it does not settle the matter: "The point of a pledge is that it binds, even when it is personally costly." He gives three options he regards as honest for a pledge-signer — explain how staying is consistent with the pledge, explain why you no longer hold it, or quit — and rules out a fourth: "Wearing the pledge while saying nothing isn't one of them."
Against the argument for staying to steer, he offers a test rather than a rebuttal: "This deal may have been the clearest red line Google's Gemini project will ever face, and yet the deal came out with no concessions to ethics-concerned employees. If their influence couldn't produce a single binding provision in that situation, then when would it?" And he describes a compounding cost to credibility: a signer who stays while their company sells unrestricted AI to a military that wants exactly what the pledge forbids "teach[es] every counterparty a lesson: these safety people will not act, even at their own brightest line. The next commitment they make is worth less. Eventually, it's worth nothing."
The external silence
Turner extends the same charge outside Google. He names Stuart Russell, "a famous AI researcher who spent over a decade crusading against autonomous weapons," in whose lab Turner says he worked for years, and reports that Russell agreed onstage at a conference to push the International Association for Safe and Ethical AI to issue a statement supporting AI providers against government coercion and promised a member poll — and that "the statement and poll never materialized."
He notes that Yoshua Bengio and Geoffrey Hinton had signed the Global Call for AI Red Lines months earlier, which named lethal autonomous weapons specifically, and argues that "when standing up for red lines meant standing against a specific powerful entity (the US government), the luminaries stayed publicly silent."
The governance assessment
Turner frames Google DeepMind as "an important experiment in governance" and Hassabis's strategy as the test case. He quotes Hassabis's account to his biographer that while he initially "wanted to ensure safety in a way that would be trustless," he instead put energy "into creating real trust — trust that was actually useful," seeking a seat "at the table," "so when a safety issue comes up, [he] can help to decide it."
His assessment is comparative rather than absolute: "Hassabis has been at the table for every contract in this essay. The classified deal made zero binding concessions to the employees raising ethical concerns. Maybe his presence averted something worse, but Google's terms were near the floor of what I imagined possible. If his seat were worth what he says, you'd expect more to show for it."
The essay's conclusion generalizes from the case to a design claim: "society cannot rely on ethics-motivated people standing firm. They just don't do it reliably. We need structures: binding contracts, independent auditors, and review bodies that cannot be quietly dissolved. Eventually, we need legislation." He closes by acknowledging the difficulty — "I know hoping for a law is a painful endeavor" — while holding the negative claim regardless: "we should at least stop telling ourselves that a seat at the table works."
Turner states he has not joined another AI company and was unemployed at the time of writing.
Provenance and limits
Retrieved from transformernews.ai, the publication's canonical host, dated July 21, 2026, with the author identified in the byline and closing note. The gap-identifier verified it on 2026-07-22 and confirmed that the "all lawful use" demand, the Anthropic "supply chain risk" designation, the OpenAI-loopholes claim, and the Google fewer-restrictions claim match material already folded onto Google DeepMind from the same URL.
Three limits bear on use. The essay is a resigning employee's account of a dispute with his former employer, and Google's side is not represented in it. Several of its load-bearing claims are explicitly Turner's inference rather than observation — most directly, his surmise about whether Dean threatened to resign, which he flags as a guess. And the Google contract is classified, so the comparative claim that its restrictions are fewer than OpenAI's is reported rather than verifiable from the documents.
Relationships
- supports: Google DeepMind — participant account of the internal opposition to the April 2026 classified deal
- related: Anthropic v. United States (Pentagon ban challenge) — the "supply chain risk" designation Turner cites as the cost of refusal
- related: Autonomous Weapons, DOD — Department of Defense (AI Deployer), Frontier AI Governance
- related: Alex Turner (author), Demis Hassabis, Jeff Dean, Stuart Russell