Google DeepMind Engineer Resigns in Protest Over Military AI: The Governance Vacuum Behind a $1.2 Billion Contract

On October 21, 2026, Google DeepMind research engineer Sofia Liguori spoke to the BBC and announced her resignation in protest of Google’s AI contracts with the U.S. Department of Defense and the Israeli military, totaling more than $1.2 billion. Earlier that month, AI researcher Andreas Kirsch had already left first, and both cited the same reason: fundamental flaws in Google’s governance mechanisms for deploying AI in military and classified environments.

The Core of the Allegation: Not Opposition to War, but Opposition to Loss of Control

Kirsch’s resignation statement was precisely worded: “Google’s governance and oversight mechanisms are insufficient to deploy AI safely and responsibly in military and classified environments.” The meaning of this sentence is technical, not moral preaching. He did not say AI should not be used for military purposes; he said existing control processes cannot keep up with the speed and scenarios in which AI is actually invoked.

Liguori’s allegation is more personal, but it likewise points to a structural problem. She joined DeepMind in 2023, researching power grid optimization—a purely civilian engineering problem. According to The Intercept, on the day after Google publicly announced in April of this year that it had signed a classified AI contract with the Pentagon, she stopped working by taking sick leave and annual leave. In her internal resignation letter, she wrote: “I no longer believe our technology will have a net positive impact on society.”

Both acknowledge that they are bound by the non-disparagement clause in their employment contracts. Liguori said directly: “In principle, I am barred for life from publicly criticizing Google.” Nevertheless, she still chose to speak to the BBC, on the grounds that the public has the right to know the true situation of military AI deployment. Choosing to speak at the edge of the legal red line is itself a real-world test of the space for internal dissent at the company.

Google’s Response, and the Questions It Did Not Answer

A Google spokesperson said the company reviews all public-sector contracts to ensure they comply with its “Acceptable Use Policy” and “AI Principles,” and stressed that the company provides internal reporting channels and “strictly prohibits retaliation of any kind.” The statement also specifically noted that the Pentagon contract explicitly prohibits autonomous weapons and unsupervised mass surveillance.

But this response did not touch the core question raised by Kirsch: in a classified environment, how can external reviewers verify whether AI use actually complies with these principles? The AI Principles are a public document; the contract is confidential. Only Google knows where the two intersect.

This is precisely the essence of the governance vacuum: the commitments are public, the implementation is opaque, and the oversight mechanism is self-declared.

Historical Precedent: Project Maven in 2018 and What Followed

This is not the first time Google has faced internal protests over military AI. In 2018, Google participated in the U.S. Department of Defense’s Project Maven, providing AI technology for drone image analysis. About 4,000 employees signed a petition explicitly demanding that the company withdraw, and eventually more than ten employees resigned. Under enormous pressure, Google announced in June of that year that it would not renew the Maven contract.

Compared with 2018, the number of resignations this time is clearly smaller, with only two people speaking publicly. But the scale of the contracts is entirely different: Project Nimbus is a $1.2 billion cloud computing and AI services contract with the Israeli government that Google won jointly with Amazon, signed in 2021; according to a 2024 report by Time magazine, its potential scale could exceed $3 billion. The Pentagon contract amount is separate and involves classified projects.

In other words, the contracts are larger, but the voices of internal resistance are smaller. This contrast itself says something.

Has the Protest Failed? Or Has It Changed Form?

The logic in 2018 was: collective action could impose sufficient costs for Google to weigh and withdraw. The result did work once. But that was in an era when both AI capabilities and the scale of defense procurement were smaller.

According to the BBC, citing an anonymous DeepMind employee, although some employees are leaving, Google’s new recruitment pipeline can fill the vacancies, and the internal culture may gradually shift over time. This sentence captures the cold logic of reality: when labor supply is ample and contract profits are large enough, the economic pressure that a few resignations can create is close to zero.

This means the function of such protests has changed—no longer a bargaining chip capable of forcing the company to change its decisions, but more an act of signaling to the public: some people inside believe this matter requires external oversight.

Liguori knew she could not publicly criticize Google, yet still chose to speak to the BBC. This is precisely the practice of that logic—pushing the issue into public view and letting regulators, researchers, and users judge for themselves.

The Real Question: Can AI Principles Constrain Classified Contracts?

Google published its AI Principles in 2018, including a commitment not to develop weapons or technologies intended to directly harm people. After that, it signed Project Nimbus, and then the classified Pentagon contract. Each time, the official position was that “the contract complies with the principles.”

But whether the principles can actually constrain classified contracts cannot be verified by any external body. This is not a problem unique to Google, but an institutional challenge facing the entire AI industry: a self-declared ethical framework naturally fails in the face of classified government procurement, because transparency is its only enforcement mechanism.

When Kirsch says “inadequate governance,” he is essentially saying: your principles are written for the public to see, not meant to actually operate in classified environments. Google’s statement did not directly respond to this point, and it is difficult for it to do so.

The resignations of the two engineers will not change the direction of the $1.2 billion contract. But the value of this event lies in the clear chain of evidence it provides: named employees, specific contract amounts, and explicit governance allegations—placing the issue on the table more directly than any academic paper.

The real risk of military AI is not a robot rebellion from the movies, but this kind of scenario in reality: a set of principles written for the public is quietly bypassed in the name of “compliance with the principles” in classified settings that cannot be audited. Liguori and Kirsch chose to speak out not to win, but to leave a record of this matter.