"Ashamed to Work at Google": Two DeepMind Researchers Quit Over Military AI Contracts, BBC Reveals Deal Details

On October 21, 2026, BBC Eye published an investigation with a blunt headline: "Google DeepMind researchers quit over military AI contracts." At its core are two named sources — not anonymous leaks, not a leaked internal memo, but two people who have already left, giving their names on the record to the BBC's cameras and recorders. One of them, London-based research engineer Sofia Liguori, delivered the line the BBC chose as the story's hook: "Realising my company is pushing more and more for AI systems in military operations makes me feel ashamed to work at Google."

In the public debate over AI ethics, "quitting" is a rare move. Protest letters can be co-signed, social media posts can be shared, questions can be asked at all-hands meetings — but resignation puts your own career on the line, and it is irreversible. This piece takes the story apart: who quit, what contract details the BBC uncovered, how this timeline runs from 2018's Project Maven to today, and what it means for those of us who call model APIs every day.

1. Who quit: two people, two different kinds of goodbye

Sofia Liguori, a research engineer who spent three years at DeepMind's London office, quit this week. According to the BBC, she agreed to be interviewed knowing full well that her contract restricts her from talking about the company publicly. Another of her quotes deserves to be reproduced in full: "I don't agree with this direction and this has led me to lose faith in my company being true to its stated ethical values." Note her wording — she is not attacking a single contract, but the company's "stated ethical values," the ethics it wrote down in black and white. This is the classic logic of a principle-driven resignation: I am not leaving because the company went bad, but because it betrayed what it wrote down itself.

The second person is AI researcher Andreas Kirsch, who left in early October. His angle is more technical, focused on governance. He told the BBC that Google's governance and oversight are "not sufficient or meaningful enough to deploy AI safely and responsibly in military and classified settings." His sharpest point was about control: "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."

Read together, the two statements form a complete critical framework: Liguori speaks of "betrayed values," Kirsch of "failed governance." One addresses what ought to be, the other what is; one says "the company shouldn't do this," the other says "even if it wanted to, it couldn't manage it."

Liguori's resignation comes with a companion farewell letter, "Stay Human," published October 20 on her own website (stay-human.sofialiguori.com) — her last day at DeepMind. The letter discloses two key facts: first, she had been on leave since April, when Google signed its classified contract with the Pentagon — her decision to leave was not impulsive, but six months in the making; second, over the past year she had been campaigning inside the company against Google's work with Israel under Project Nimbus, with Lockheed Martin, and with US immigration and border agencies. In other words, she walked the "internal opposition" path first, found it blocked, and only then chose to leave and speak publicly.

The letter also contains a striking piece of self-positioning: "I don't think resigning is an effective method, certainly not as an individual act." She does not call on colleagues to follow her out. Instead she asks those who stay to join the GDM Union — UK staff organizing through the UTAW and Unite unions — whose demands include enforceable red lines on military and surveillance applications, more transparency around contracts, and the right to abstain on ethical grounds. She defines resignation as the end of a personal choice, and organized action as the real starting point. It is a clear-eyed judgment: an individual resignation is a moral statement; collective bargaining is power.

The BBC notes this is not the first principled exit from DeepMind this autumn: in September, engineer Robert O'Callahan quit because his team was making AI faster, and two current employees told the BBC anonymously that "a lot of people have left the company that feel strongly about this." There is no exact number for the scale of the exodus, but the structure — two named, anonymous corroboration, a third case in autumn — suggests this is not an isolated incident.

2. The contract details the BBC uncovered: the killer phrase is "any lawful government purpose"

The hard core of the reporting is the contract language. As quoted by the BBC, Google's agreement with the Pentagon lets its models be used for "any lawful government purpose," with a single restriction: they "should not be used for domestic mass surveillance or autonomous weapons (including target selection) without appropriate human oversight and control."

Unpack that sentence and it has three layers of problems. First, in "any lawful government purpose," the power to define "lawful" sits with the government itself — a self-referential clause that hands the judgment of use entirely to the user. Second, the restriction smuggles in a giant back door: "without appropriate human oversight and control" — which is to say, with "appropriate human oversight," autonomous weapons and mass surveillance are fine. And who defines "appropriate"? The contract doesn't say. Third, Kirsch's criticism lands exactly here: Google has "no right to control or veto," and "limited visibility" — it signed, it got paid, but where the models go and how they are used after handover is beyond its own sight.

In her farewell letter, Liguori calls this kind of contract design "a very convenient contract design for executives to escape accountability": it allows AI to be built into military and surveillance systems "while imposing no restrictions on permissible uses and leaving the company no monitoring power." Her judgment: this is not negligence, it is design — designed so that the signing party can say "we didn't know" and "we couldn't control it" when things go wrong.

Google's response is duly recorded by the BBC. A spokesperson said the company works with public-sector organizations worldwide under "a consistent process for reviewing and entering these contracts," in compliance with its acceptable use policies and AI principles; that it has grievance channels; and that it has "strict workplace policies in place to prevent retaliation of any kind." It is the standard corporate response template: no denial that the contract exists, no discussion of specific terms, everything folded into "process compliance." Notably, Google did not respond to the specific allegations in Liguori's letter (such as staff being called into HR meetings for raising military contracts), and the BBC notes explicitly: "Google hasn't commented on the letter."

The reporting also widens the lens: Google is not alone in signing such military AI deals this year. The list includes OpenAI, Amazon, Microsoft, SpaceX, Oracle, Nvidia, and the start-up Reflection. Anthropic is the counterexample: its Claude was dropped from Palantir's Maven Smart System, the Pentagon's main targeting platform, after Anthropic refused to relax its limits on military use. The contrast carries real information — "sign" and "don't sign" are both genuine options; Anthropic proved refusal is possible, at the cost of losing the order. The backdrop: in January, US Defense Secretary Pete Hegseth declared the US would become an "AI-first warfighting force," the Pentagon has since set up a command for autonomous warfare and begun bulk procurement of AI agents that can act on their own. The demand-side engine is running at full speed; every lab on the supply side is answering the same multiple-choice question.

3. An eight-year timeline: in 2018 employees pulled Google out; in 2026 employees walk out themselves

In 2018, thousands of Google employees protested the company's work on Project Maven, the Pentagon effort to analyze drone footage with AI. Some resigned, Google let the contract lapse, and it published the first version of its AI Principles — which explicitly listed AI applications "we will not pursue," including weapons and surveillance that violate international norms. That was a victory for collective employee action: the company was pulled out of military AI.

In 2024, the $1.2 billion Project Nimbus cloud contract with the Israeli government, shared with Amazon, drew a new round of staff protests. In February 2025, Google rewrote its AI Principles, dropping the explicit "will not pursue" list — the rewrite announcement was co-signed by DeepMind chief Demis Hassabis, citing "a global competition taking place for AI leadership." In April 2026, Google signed the classified Pentagon contract. In October 2026, two researchers resigned and the BBC Eye investigation landed.

Worth a word on Nimbus 2024 on its own: it was a $1.2 billion cloud deal, carried jointly by Google and Amazon, serving the Israeli government. The employee protests it triggered never reached the scale of 2018's Maven protests, but it was a key transitional signal — from "helping the Pentagon watch drone footage" (Maven, analytical AI), to "providing a government with full cloud infrastructure" (Nimbus, platform partnership), to "delivering models into classified systems" in 2026 (capability delivery). The form of cooperation kept escalating while employees' bargaining power kept falling: in the Maven era staff could still get a contract "allowed to lapse"; in the Nimbus era protests did not stop the deal; by 2026 even "what the contract says" had become classified. Three jumps, one direction.

Put 2018 and 2026 side by side: 2018 was "employees stay, the company turns"; 2026 is "the company stays, employees turn." Eight years, and the arrow has flipped. The 2018 protests worked because the AI Principles of the time contained a red line in black and white that employees could invoke; the 2026 resignations feel inevitable because that red line was erased by the company itself in February 2025. When Liguori says she no longer believes the company will stay true to its stated ethical values, the subtext is: after 2025, the company's stated values themselves changed — she remains loyal to the 2018 version.

This timeline deserves a second look: at every turning point, changes to the "principles" text came before changes to the "contract" reality. The text is the weathervane; the contract is the landing. When a company starts revising its principles wording, the real question isn't the wording — it's that the company is already clearing space for its next contract.

4. Inside the building: fear, HR meetings, and a retired general's verdict

The BBC piece carries another set of details — about why only two people went on the record. Liguori says her contract "in principle" bars her from criticizing Google for life: "In principle I'm not allowed to criticise Google for the duration of my life." Her answer: "talking about facts is not criticising." Inside the company, she says, "people are scared of retaliation and putting their career at risk."

The farewell letter makes the fear more concrete: employees who raise military contracts with their teams get "called into HR meetings on the pretext of having breached some internal policy" — which Liguori calls intimidation. One section of the BBC report carries an employee quote as its subhead: "You feel you are being watched." Together these explain why named resignations are so rare — the cost isn't "losing the job" (the job is already gone), but the lifetime gag-style clauses in the contract, plus reputational risk inside the industry. Liguori chose to put her name on the line precisely because anonymous leaks were no longer enough on this issue: she tried internal channels for a year, and internal channels sent her to an HR meeting.

The report also quotes a heavyweight: Jack Shanahan, the retired lieutenant general who led Project Maven — the military chief of the very project Google employees protested back then. He told the BBC the philosophy now seems to be to "go as fast as possible to deploy [AI technology] as widely as possible, potentially at the expense of the rigour and discipline of test and evaluation." A general who once stood on the "military wants AI" side now publicly criticizes deployment speed steamrolling testing rigor — the statement itself shows that even inside the demand side, people think the vehicle is moving too fast.

And the demand side is indeed accelerating: after Hegseth's January framing, the Pentagon set up an autonomous-warfare command and began procuring "AI agents that hack on their own." Note the direction of procurement — no longer "analyzing drone footage" assistive tools (the 2018 Maven shape), but agents that act autonomously. When the shopping list moves from "assisting decisions" to "making decisions," the sentence "no right to control or veto" carries a completely different weight. The 2018 debate was "should AI help the military watch footage"; the 2026 debate is "should AI decide for the military" — the problem got hotter while the seller's visibility got dimmer (classified contracts).

5. Why it matters: this isn't gossip, it's the power structure of the AI supply side being rewritten

First, an easily missed judgment: the most important information in this story isn't "someone quit," it's the contract terms themselves. "Any lawful government purpose" plus "no right to control or veto" plus "limited visibility" — together they describe a new kind of military-industrial relationship. In traditional defense contracts, the seller at least knows what it sold and who it will be used against; in contracts for a general-purpose technology like AI models, the seller hands over a "capability" whose uses are defined by the buyer in deployment. Kirsch's criticism hits exactly this structural problem: when the technology is general-purpose and the deployment is classified, the seller's ethics review ends the moment the ink dries — everything after happens out of its sight.

Second judgment: the 2018 playbook won't replay. The 2018 employees won on three conditions: social-media-era amplification of public opinion, the company at the peak of its "don't be evil" brand persona, and a red line in the AI Principles they could invoke. All three conditions are gone — public opinion has gone numb on "AI militarization" (the Pentagon openly declaring an AI-first warfighting force barely made waves), every lab is chasing government orders (seven giants on the list, not one missing), and the red-line text was deleted in 2025. So 2026's resisters can only choose the atomized resistance of individual resignation. Liguori sees it clearly herself: individual resignation is not an effective method. Her turn toward the union route is precisely because she realized that in an era with no red-line text to invoke, resistance has to return to its oldest form — collective bargaining.

Third judgment: Anthropic's counterexample shows refusal is possible, but the price is real. Claude being dropped from Palantir's Maven system means losing orders from the Pentagon's main targeting platform. That means "ethical stance" has an explicit price tag in the 2026 AI market — you can choose not to sell, but you bear the cost of losing one of the largest customers. It is a public pricing signal to every AI lab: the market has already quoted how much principles are worth.

6. What it means for vibe developers: four concrete takeaways

This is closer to those of us who call APIs and ship apps every day than it looks. First, where the model you call is being used is becoming part of your product's narrative. When your app's README says "Powered by X," X's military contracts become questions your users may ask you. This isn't moral blackmail; it's supply-chain transparency extended — just as you'd care whether your cloud provider has a breach history, a model's "usage history" is becoming a dimension of technology choice. Anthropic getting dropped from an order for refusing to relax limits will itself become some developers' reason to choose it.

Second, "explainable governance" is becoming a selling point for infrastructure. Inference frameworks talk about multi-tenant QoS; Anthropic talks about usage-policy red lines — the competitive dimensions of AI infrastructure are expanding from "how fast, how cheap" to "can you say clearly where models get used." For indie developers, this means: when choosing infrastructure, asking "how much control does this vendor have over model usage" isn't preciousness, it's risk management.

Third, the 2018 lesson still holds for small teams today: write the red line down. The reason Google's 2018 AI Principles could be wielded by employees is that they were published in black and white. The same goes for your small product — writing "our AI will not be used for X" in your README or on your site costs nothing, but at the critical moment it is the only text you can invoke. Liguori stays loyal to the 2018 text; if your product grows up one day, which paragraph that you write today do you want your early users to stay loyal to? Figure it out, write it down, publish it.

Fourth, pay attention to who negotiates on your behalf. Liguori's final choice was to join a union — converting individual moral judgment into collective bargaining chips. The indie-developer version: join or follow the governance bodies of your ecosystem. API terms of models, revisions to open-source licenses, platform content policies — changes to these "rule texts" eventually become your product's compliance cost. The reason Google's 2018 employees could win is that they were an organized collective; the reason 2026's Liguori could only resign individually is that the vehicles of collective action (unions, published principle texts) had been dismantled in advance. Rule texts and bargaining vehicles are the only two weapons ordinary people have — don't discover they're gone when you need them.

7. Honesty notes

The factual core of this piece comes from BBC Eye's October 21, 2026 investigation (verified by opening the madrobot.blog evidence-chain page, which contains links to the original BBC article and the farewell letter; the bbc.co.uk domain is blocked by this site's fetch policy, so the original page could not be opened directly), and from Sofia Liguori's October 20 farewell letter "Stay Human." Contract details are as quoted by the BBC — phrases like "any lawful government purpose" are contract language cited in the reporting, not a contract we have seen. Google's response, Anthropic's removal from the Palantir system, the 2018 Maven protests, and the 2025 AI Principles rewrite are all as recounted in the same reporting. madrobot.blog serves here only as the evidence-chain relay page, not as an independent source. The claim that "a lot of people have left" comes from two anonymous current employees cited in the report; there is no confirmed number for total departures, and this piece makes no claim about scale. Restrained by design: we relay facts and offer judgments, without stoking positions.