The resignations at Google DeepMind are not just a staffing story. They are a governance story, because they ask a question no AI company has yet answered convincingly: once a frontier model is deployed inside a classified military network, who actually controls what it does?
Andreas Kirsch, an AI researcher who resigned earlier in October, argued that Google’s governance and oversight were not sufficient or meaningful enough to deploy AI safely and responsibly in military and classified settings. He said Google seemed to have no right to control or veto how the Pentagon uses its AI, and only limited visibility into how it was being deployed.
The classified problem
Classified deployment is different from ordinary enterprise use. The customer is a government, the environment is closed, and the vendor’s usual tools for monitoring, auditing and intervening may not apply in the same way. Google’s Pentagon agreement allows its models to be used for any lawful government purpose, with stated restrictions on domestic mass surveillance and autonomous weapons without appropriate human oversight and control.
Kirsch’s argument is that those restrictions are only as strong as the oversight behind them. If the vendor cannot see how the system is used, cannot veto a particular use, and cannot verify that human oversight is actually in place, then the restrictions are promises rather than controls. Sofia Liguori, a research engineer who resigned after three years, made the personal version of the same point, saying the push into military operations made her feel ashamed to work at Google.
What Google says
Google said it reviews public sector contracts for compliance with its acceptable use policies and AI principles, and that employees have dedicated channels to raise ethical concerns without retaliation. The company also said its military agreements include restrictions on mass surveillance and autonomous weapons.
As the Financial Express reported, the resignations follow months of internal opposition, including a petition signed by around 600 employees in April urging leadership to stay away from classified military work.
The governance question will not be settled by a resignation letter, but the letters have forced it into the open. For any company selling AI into classified settings, the next test is not a policy document. It is a credible answer to Kirsch’s question: show the oversight, show the veto, and show the visibility, or accept that the people who build the systems will keep walking out.
