The BBC has reported that several current and former Google DeepMind employees have resigned, or are leaving, over the company's AI contracts with the military. The story, by Christopher Giles of BBC Eye investigations and published on October 21, 2026, centers on research engineer Sofia Liguori and AI researcher Andreas Kirsch, who both told the BBC they lost faith in Google's governance of military uses of its models. Read the original at the BBC.
This is a labor and policy story with real technical stakes: how much human oversight remains when AI helps pick targets under time pressure, and how much control a model maker keeps once a government deploys its system in classified settings. Below we separate what is on the record from what is claimed, add the contract background from earlier reporting, and list what to watch. It sits beside our coverage of Anthropic's Pentagon fight and the Pentagon AI overreliance strike probe.
TL;DR: claimed vs verified
| Question | Answer | Status |
|---|---|---|
| Did DeepMind staff resign over military work? | Liguori (this week) and Kirsch (earlier in October) say so | On the record to the BBC |
| What does the contract allow? | Classified military use, "any lawful government purpose" | BBC; earlier reports on the April signing |
| What are the stated limits? | Not for domestic mass surveillance or autonomous weapons, including target selection, without appropriate human oversight and control | BBC reading of the agreement |
| Can Google veto a use? | Kirsch says Google seems to have no right to control or veto the Pentagon's use | One departing researcher's view |
| Lifetime non-disparagement? | Liguori says she cannot criticise Google for her lifetime | Employee account, Google not quoted on it |
| How many staff left? | "Several"; no total published | Unquantified |
| Google's position | Contracts follow its review process, acceptable use policies and AI principles | Google spokesperson to the BBC |
What the resigning staff said
Liguori, a London-based research engineer, resigned this week after three years and spoke to the BBC in a personal capacity despite what she described as contractual limits. "Realising my company is pushing more and more for AI systems in military operations makes me feel ashamed to work at Google," she said. She tied her decision to Google's increasing work with the US and Israeli militaries and said it led her to lose faith in the company being true to its stated ethical values.
Kirsch, who resigned earlier this month, told the BBC he believes "Google's governance and oversight are not sufficient or meaningful enough to deploy AI safely and responsibly in military and classified settings," or to develop superintelligence safely in future. He added that Google "seems to have no right to control or veto how the Pentagon uses its AI and also seems to have limited visibility into how it's being deployed." He also said current models are not robust enough for such tasks and have not been trained for them.
An anonymous current DeepMind employee said some motivated staff who joined for the original mission could leave but that a pipeline of early-career people could fill the roles, calling it potentially "a changing of the guard." Another said many who felt strongly have already left.
What the contract reportedly says
The BBC says Google is one of several companies that signed contracts with the US military earlier this year to deploy AI across classified environments. The agreement says Google's models can be used for "any lawful government purpose" and "should not be used for domestic mass surveillance or autonomous weapons (including target selection) without appropriate human oversight and control."
Earlier reporting around the April signing, summarized by outlets including The Decoder, described an employee letter to CEO Sundar Pichai signed by more than 600 staff, including senior leaders, urging the company to reject classified work because classification hides how the technology is used. Those reports say Google Public Sector described the deal as extending an agreement from November, and that the contract language reportedly also says it grants no right to control or veto lawful government operational decisions. Legal experts cited in that coverage said the safety clauses were not legally binding. We have not read the contract itself, so these are secondary accounts.
Why "should not" matters
The difference between "must not" and "should not" is the heart of the dispute. A prohibition with enforcement would let Google audit or stop a use. A statement of intent, paired with a clause that denies any veto, leaves the model maker relying on the customer's good faith. The BBC says similar contracts were signed by OpenAI, Amazon, Microsoft, SpaceX, Oracle, Nvidia and Reflection, though it does not say their wording matches Google's. Anthropic, by contrast, fought the Pentagon over similar terms, as our DC Circuit post explains.
Oversight at machine speed
The BBC also quotes outside experts on the practical problem. Retired Lieutenant General Jack Shanahan, former director of Project Maven, described a philosophy of going as fast as possible to deploy AI as widely as possible, "potentially at the expense of the rigour and discipline of test and evaluation." Sarah Shoker, formerly head of OpenAI's Geopolitics team, said AI compresses the time an operator has to decide, and that with hundreds of targets chosen by AI in a short time it is almost impossible for a human to meaningfully review outputs for errors. The Pentagon's line, in the words of under secretary Emil Michael, is that "speed defines victory in the AI era."
These points connect to the research on automation bias and to incidents we have covered, such as US military AI producing false intelligence about a Chinese ship. The claim that a human is "in the loop" is only meaningful if the human has time, information and authority to say no.
Google's response and the gag-clause claim
Google told the BBC it works with public sector organisations worldwide and "follow[s] a consistent process for reviewing and entering these contracts, including compliance with our acceptable use policies and AI principles." It added that it has dedicated channels for employees to raise concerns and strict workplace policies against retaliation. The BBC notes that employees said staff are afraid to speak up because contracts prohibit publicly criticising the company, and Liguori said people fear retaliation and feel watched.
We could not verify the contract terms or any lifetime clause independently. What is established is that a named engineer made the claim on the record about her own contract and that Google's statement addresses retaliation but not that specific term.
Context: Google's history with military work
The BBC recalls the 2024 controversy over Project Nimbus, the $1.2 billion cloud and AI contract with Amazon for the Israeli government and military, which Google said at the time was not directed at highly sensitive, classified or military workloads relevant to weapons or intelligence. One anonymous DeepMind employee told the BBC that many people who felt strongly have left as the company changed its ethics position, which is the staff's characterization rather than a documented policy timeline. The BBC also notes that Israel's AI capabilities in warfare have been debated, including in a documentary examining AI-powered targeting in Gaza; an Israeli government minister has said the claims in the film were false and distorted, so that debate remains contested. For readers tracking the lab-level debate, compare how other departures have been framed in our posts on Robert O'Callahan's DeepMind exit and OpenAI safety lead David Robinson quitting; the DeepMind resignations here are about deployment policy, not research pace.
What people are asking
Is Google leaving the military market? Nothing in the reporting suggests so. Google's statement defends its contract review process and refers to its AI principles, and the company has not announced any change.
Does this affect Gemini users? Not directly. The coverage concerns government deployments, not consumer or enterprise access to the models.
What to watch next
- A count. The BBC says "several." A number of departures, or a collective letter, would show scale.
- Contract disclosure. Whether Congress, a court or Google itself publishes the operative language.
- Policy changes. Any revision to Google's AI principles or acceptable use policy on classified work.
- Peers. Whether employees at other signatories speak up, and whether any company adds an enforceable human-control clause.
- Retaliation claims. Google says it prohibits them; any formal complaint would test that.
What this means for people who build with these models
If you use Google's models through an API, nothing in the reporting changes your access. The practical takeaway is about vendor risk and values: your provider's policies on government use are now part of the supply chain you can evaluate, alongside licensing and data handling. Teams with customers who care about military use may need to document which model providers they rely on. Our explainer on AI policy in 2026 shows how one lab frames its own commitments publicly.
Related reading
- DC Circuit overturns Anthropic Pentagon blacklist ruling
- Pentagon AI overreliance strike and Senate probe
- US military AI and false intelligence about a Chinese ship
- Hegseth autonomous warfare panel and protesters
- Robert O'Callahan leaves Google DeepMind
- OpenAI's David Robinson quits
- Anthropic Public Benefit report
Quotes and details are as reported by the BBC on October 21, 2026 and in earlier outlet coverage; contract terms are not independently verified.