Sofia Liguori, a London-based research engineer, resigned this week, while AI researcher Andreas Kirsch said he also left earlier this month because he believed Google’s governance and oversight were inadequate for military deployments.
The Pentagon agreement allows Google’s models to be used for “any lawful government purpose.” It says the systems should not be used for domestic mass surveillance or autonomous weapons, including target selection, without appropriate human oversight and control. Employees and outside experts questioned whether people could meaningfully review AI outputs when systems operate rapidly and generate large volumes of recommendations in high-pressure settings.
Google says it reviews public-sector contracts under its acceptable-use policies and AI principles, which include commitments to international law, human rights and human oversight. DeepMind chief Demis Hassabis has said the company’s values remain intact, while arguing that changing geopolitics and the broad availability of AI require technology companies to help democratic governments develop defensive capabilities. He has also said he remains opposed to lethal autonomous weapons.
That change reversed a position adopted after employee protests over Project Maven, a military program that used machine learning to interpret drone imagery. Nearly 200 DeepMind workers later signed an internal letter calling for military contracts to be dropped and for stronger governance over how the lab’s technology reaches military customers.
Google’s work with Israel has added to the pressure. Employees have criticized Project Nimbus, a $1. 2bn cloud and AI agreement with Amazon serving Israeli government ministries and the military. Google has said the arrangement is not intended for highly sensitive, classified or weapons-related workloads. The latest resignations show that the company’s shift toward defense partnerships remains a live source of conflict inside its AI division.