Google DeepMind engineer quits over military AI contracts and handling of dissent

Google’s military AI work has cost DeepMind a research engineer who says safeguards do not answer her basic objection to using AI in war. Sofia Liguori resigned this week after three years at the company, citing its work with the US and Israeli militaries, the BBC reported on October 21.

From electrical grids to internal protest

Liguori joined DeepMind in 2023 and worked on AI systems intended to route power through electrical grids more efficiently. In an interview published by The Intercept on October 21, she described DeepMind as a dream employer: a research-oriented place where computer scientists could work alongside people trained in biology, physics, mathematics and other fields.

She said she had not known about Project Nimbus when she joined. Learning about Google’s work with Israel led her to look for colleagues who shared her concerns. Initially, she struggled to find them inside a large, siloed company. A colleague later connected her with union organizers and the employee advocacy group No Tech for Apartheid.

Finding that collective effort gave her the strength to stay for at least another year, she said. Her activism included messages on corporate channels, conversations with colleagues and co-organizing a February demonstration outside the office against AI in warfare. Protesters carried banners reading “No AI in the kill chain.” She sought to turn colleagues’ dissatisfaction into collective action.

I’m against AI for war, and I don’t think a commercial company like Google should have anything to do with that.

Sofia Liguori, speaking to The Intercept

Why human oversight did not settle her objection

The military agreement at the center of her protest reaches beyond office administration. In a May 1 announcement, the Pentagon named Google among eight companies agreeing to supply AI capabilities for classified networks. It described the intended uses as supporting warfighting, intelligence and enterprise operations, including combining data and helping personnel make decisions in complex operational settings.

Google’s agreement permits its models to be used for “any lawful government purpose,” according to the BBC. It says the technology “should not be used for domestic mass surveillance or autonomous weapons (including target selection) without appropriate human oversight and control.” The wording makes human oversight a condition on those uses, rather than imposing an unconditional ban.

Liguori’s objection is broader. She argues that AI targeting can reduce people to aggregated data points while obscuring how a target was defined. A person formally approving an output may not be able to meaningfully evaluate it, she said. In her view, responsibility can then become dispersed, allowing everyone involved to say they did not make the decision.

She also questioned the focus on model errors alone. Although she is concerned about reliability, she told The Intercept that treating errors as the central problem risks implying automated targeting would be acceptable if it worked correctly. She acknowledged that military AI was debated inside DeepMind, with some colleagues supporting it on national-defense grounds.

A dispute over how dissent is heard

Liguori said her attempts to challenge the work internally met procedural barriers rather than substantive answers. She alleged that some posts were deleted because they contained the word “genocide,” which she was told violated internal rules. She also said managers and HR criticized her tone and directed her toward other channels for raising concerns.

Those channels, she said, produced semi-automated responses without meaningful engagement. What she wanted was discussion and clear answers from leadership. Looking back on the organizing effort, she said petitions had limited effect and argued that employees should have stopped working to try to prevent the military contract.

Asked about her criticisms, Google said it follows a consistent review process for public-sector contracts, including compliance with its acceptable-use policies and AI principles. The company also said it has dedicated channels for employees to safely raise concerns ranging from business ethics to misconduct, and strict workplace policies intended to prevent retaliation of any kind.