Alex Turner, a Google DeepMind researcher, resigned after months of internal efforts to block a Google contract that gives the Department of War broad access to the company’s AI models, according to an account Turner published and relayed by Zvi Mowshowitz on his Substack newsletter, Don’t Worry About the Vase. The contract, in Turner’s telling, carries no restriction against use in autonomous weapons or mass surveillance. Turner is, so far, the only Google DeepMind employee to leave over the decision.
The stakes trace back to 2018, when DeepMind and several of its senior leaders, including CEO Demis Hassabis, chief AGI scientist Shane Legg, and Google chief scientist Jeff Dean, signed a public pledge not to develop or support lethal autonomous weapons. Google’s AI Principles carried a matching prohibition on weapons and surveillance use until 2025, when the exclusion was dropped, according to DeepMind researcher Andreas Kirsch’s timeline of the policy’s history as cited by Mowshowitz. The 2026 contract followed the next year.
Turner describes trying several channels before resigning. He lobbied the International Association for Safe and Ethical AI to issue a statement backing Anthropic, which had separately pushed back against Pentagon pressure. A member vote reportedly favored the statement, but the organization never issued one. Inside Google, Turner circulated a petition to Dean that drew more than 250 signatures and proposed Dean lead a review body. Dean signed an amicus brief supporting Anthropic but, in Turner’s account, did not use his full leverage to stop the deal. Turner says a direct appeal to Hassabis was routed to policy staff and left unresolved until Google signed the contract. Separately, more than 600 DeepMind employees signed an open letter opposing it.
Mowshowitz, writing as commentary rather than reported fact, places more responsibility on Hassabis than on Dean, arguing that Hassabis keeps describing DeepMind’s principles as intact even after co-authoring the policy update that dropped the weapons restriction. Hassabis has said the company’s commitments are unchanged; Mowshowitz contends the two positions cannot both be true. That judgment belongs to Mowshowitz. Hassabis has not addressed Turner’s resignation directly in public.
The resignation surfaced days after Hassabis published a separate essay proposing a Frontier AI Standards Body, modeled loosely on the financial regulator FINRA, under which the biggest AI developers would let a reviewing body examine models on a voluntary basis, with up to 30 days of lead time before launch. Hassabis has framed the body as adaptable enough to coordinate a slowdown across labs if conditions warranted it. Critics quoted by Mowshowitz, including MIRI researcher Nate Soares and AI policy analyst Aaron Scher, called the plan voluntary, weighted toward public releases over internal deployment, and vague about what happens once a model is judged dangerous.
The juxtaposition matters for anyone who treats a frontier lab’s safety pledges as a signal of how it will act under commercial or government pressure. DeepMind’s 2018 weapons pledge, reaffirmed for years, did not survive one defense contract negotiation, despite objections from hundreds of employees and pressure from its own chief scientist. Operators evaluating vendor or partnership ties to labs pursuing government and defense work should treat published principles as aspirational until a lab has faced a comparable test and held the line.
Zvi Mowshowitz reported and analyzed these events in a July 2026 post, “Demis Hassabis on the New Coming Age,” on his Substack newsletter Don’t Worry About the Vase.
