AI Researcher Quits DeepMind After Google Signs Unrestricted Pentagon Deal

A Google DeepMind research engineer describes an internal campaign that ended with his resignation after Google signed a military AI contract with the Pentagon that carried no restrictions against autonomous weapons or mass surveillance. His effort began as a push to get Google to divest from DHS and ICE work — citing Google Cloud contracts reaching immigration agencies through third parties, the removal of apps that flagged ICE activity, and the handover of a student protester’s account to ICE in apparent violation of Google’s own terms — and broadened into an attempt to stop the company from accepting a Pentagon deal with no guardrails. He argues the fight was really about whether AI ethics commitments survive real pressure, and concludes they largely did not.

Rather than organize a mass walkout, which he judged Google could absorb, he targeted a handful of hard-to-replace insiders, betting that even one high-leverage departure would register with leadership. He got Chief Scientist Jeff Dean — a 2018 signatory of the pledge against killer robots — to co-sign an amicus brief backing Anthropic against Pentagon coercion, and drafted a 25-page counterproposal with concrete contract language and oversight mechanisms that surveillance-law experts endorsed. He routed it to CEO Demis Hassabis, who passed it to policy staff where it stalled. Prominent AI-safety figures he approached, including Stuart Russell, made supportive gestures that quietly evaporated.

Senior management assured him Google would not sign; it did, with terms he says were even weaker than OpenAI’s, and Dean stayed on despite his pledge. The account is notable less as a single resignation than as a case study in institutional failure: it shows how firms and individuals with public reputations for caring about AI ethics can decline, one after another, to spend their credibility when a lucrative government contract and political pressure are on the line — a warning for anyone counting on voluntary AI principles to hold under stress.