Google DeepMind Resigner: 'Can't Criticise Google for Life'
A researcher walked out of Google DeepMind this week and then did the one thing her contract says she may never do — for the rest of her life. Sofia Liguori, a London-based research engineer who spent three years at the company, resigned in protest at Google's expanding military AI work and then gave the BBC an on-the-record interview in which she said, "In principle I'm not allowed to criticise Google for the duration of my life." She spoke anyway.
That sentence is the sharpest thing anyone has said about Google in years. Not a leaked memo. Not a lawsuit. A departing engineer, named and on the record, breaking what amounts to a lifetime gag clause because — as she put it — the public should know what is happening inside the company she is leaving behind.
Liguori is not alone. The BBC reported on 21 October that several tech workers at DeepMind have resigned or are in the process of resigning over the same issue — including AI researcher Andreas Kirsch, who quit earlier this month. Their target: Google's Pentagon AI contracts, signed earlier this year, which permit the US military to use Google's AI models for any lawful government purpose.
She quit. Then she broke her lifetime silence clause.
Liguori, a research engineer, resigned in the week of 19–21 October after three years at DeepMind, where she joined in 2023 to work on AI for electrical-grid routing — about as far from weapons research as a machine-learning job gets. She told the BBC, speaking exclusively and in a personal capacity, that she had decided to leave over Google's increasing work with the US and Israeli militaries. She no longer believed, she said, that the company was true to its stated ethical values.
Then came the line that will follow her around the industry. Her contract, she told the BBC, restricts public criticism of Google for life. She chose to speak regardless, arguing that stating facts is not criticism. She described a workplace where, in her words, people are scared of retaliation, feel watched, and cannot speak. For a company that runs "dedicated channels for employees to safely raise concerns," that is a devastating sentence to have uttered in public.
The detail that makes her account hard to dismiss: she stopped working months ago. In an interview with The Intercept on 21 October, Liguori said she staged a personal "sick-out" — combining sick leave and holiday time — after Google announced its classified military AI contract with the Pentagon in April, and told the outlet she no longer believed the company's technology would have a net positive impact on society. That is not a heat-of-the-moment exit. That is six months of deliberation ending in a resignation letter.
Her path to the door began with Project Nimbus, Google's contract supplying Israel's military, which pushed her to find colleagues with shared concerns. In February she co-organised a demonstration outside the office with the campaign group No Tech for Apartheid, carrying banners opposing AI's role in the kill chain. A Google spokesperson, Tom Hewett, told The Intercept that Google works with public sector organisations under a consistent review process and enforces strict anti-retaliation policies.
The governance warning Google won't answer
Andreas Kirsch, an AI researcher, resigned from DeepMind earlier in October over the same concerns, and his warning is arguably the more technical — and more damning. "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 told the BBC. He added that 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 artificial superintelligence safely in the future.
That is the crux. The contract's fine print can say anything; once the models leave Google's hands, the company has — by Kirsch's account — neither visibility nor veto. Sarah Shoker, formerly head of OpenAI's Geopolitics team, spelled out to the BBC why that gap matters in practice: AI compresses the time an operator has to make a decision, which is especially dangerous in a high-stress environment, and when hundreds of targets are selected by AI in a short window it is nearly impossible for a human operator to meaningfully review the outputs for errors.
Google's own AI Principles promise AI implemented with human oversight and aligned with international law and human rights. But staff and experts alike told the BBC they doubt the level of human oversight "at critical moments" — the moments that are, by definition, the only ones that count.
The contract says "any lawful government purpose"; the guardrail says "should not"
Read the two sentences side by side. Google's Pentagon contract, signed earlier in 2026, permits its AI models to be used for any lawful government purpose. The guardrail language says models should not support domestic mass surveillance or autonomous weapons, including target selection, without appropriate human oversight and control. One is a binding permission; the other is a polite suggestion. Lawyers know the difference. So does everyone who has ever read a terms-of-service agreement.
Retired Lieutenant General Jack Shanahan, the former director of Project Maven who now speaks candidly about military AI, described the prevailing philosophy to the BBC as a rush: deploy the technology as fast and as widely as possible, potentially at the expense of the rigour and discipline of test and evaluation. When speed is the doctrine, "should not" is not much of a brake.
Google has been here before. In 2018 its workers protested Project Maven, the Pentagon's drone-video AI project; Google declined to renew that contract and published its AI Principles — though it has since quietly dropped the passage pledging to avoid potentially harmful weapons-related applications. The principles survived. The pledge did not.
Google's answer: process, policies and the retaliation line
Google's spokesperson told the BBC the company works with public sector organisations worldwide and follows a consistent process for reviewing and entering contracts, including compliance with its acceptable use policies and AI principles. The company, the spokesperson added, has dedicated channels for employees to safely raise concerns — from business ethics to misconduct — and strict workplace policies to prevent retaliation of any kind.
Tom Hewett, another Google spokesperson, gave The Intercept the same substance: consistent process, compliance with policies, strict anti-retaliation rules.
We checked the syndication relays ourselves: the BBC's report was carried verbatim by the Nile Post, Ada Derana and KahawaTungu, each page stamped 21 October — consistent with a same-day exclusive, not a slowly recycled wire story. That matters because Google's response reads like a template; the resignations, by contrast, have names, dates and broken silence clauses attached.
The 600-signature letter Google ignored in April
Keep the two events distinct, because Google would like you to blur them. On 27 April 2026 — six months before Liguori's resignation — more than 600 Google employees delivered an open letter to chief executive Sundar Pichai demanding the company reject a proposed Pentagon deal that would allow its AI to be deployed in classified military operations. The signatories came from Google DeepMind, Cloud and other divisions, and included more than 20 directors, senior directors and vice-presidents. The Washington Post saw a copy of the letter; Bloomberg Law first reported it.
The letter's core demand was categorical: reject classified workloads, because that was the only way to guarantee Google would not become associated with such harms — otherwise, they warned, such uses could occur without the employees' knowledge or power to stop them. It named the harms explicitly: lethal autonomous weapons and mass surveillance.
According to AFP reporting, letter organisers said Google had proposed contractual language to keep Gemini away from domestic mass surveillance or autonomous weapons without appropriate human control, while the Pentagon pushed for broad "all lawful uses" wording. The employees called the safeguards technically unenforceable, pointing to Pentagon policy that bars outside entities from controlling its AI systems. Sound familiar? It should: Kirsch is now saying the same thing on the record, alone, on his way out.
Read the April letter against the UN's new military AI forum and the pattern sharpens: institutions everywhere are trying to draw lines around autonomous weapons while the world's biggest AI labs keep signing contracts that cross them.
Hegseth's AI-first warfighting force — and the rush to sign
None of this is happening in a policy vacuum. US Defence Secretary Pete Hegseth declared in January 2026 that America would become an AI-first warfighting force, and AI-first military integration is a defining priority of the Trump administration — see our explainer on Trump's AI-first military push for the full picture of how Washington got here.
And Google is far from alone. Similar AI contracts with the US military were signed earlier this year by OpenAI, Amazon, Microsoft, SpaceX, Oracle, Nvidia and the startup Reflection, covering AI across transport, planning and battlefield operations. When every major lab signs, each can claim it had no choice — a competitive pressure the industry is now deploying against its own principles, in a year that also saw the White House ending AI self-regulation.
Update, Oct. 22: the deleted posts, the union vote and the eight-company club
Liguori told The Intercept that several of her internal posts were deleted for containing the word "genocide," and that she was told "genocide is an intrinsically violent word," violating internal policy, according to AI Weekly's re-report of The Intercept's interview. Our post covers the sick-out from the same interview, but not this moderation detail. DeepMind employees voted to unionise in May, reported by AI Weekly. And the BBC named eight firms with comparable Pentagon AI deals: OpenAI, Amazon, Microsoft, SpaceX, Oracle, Nvidia and Reflection alongside Google, via FourWeekMBA's re-report of the BBC. The 2024 cloud-and-AI contract with Israel known as Project Nimbus was worth $1.2B, per TechFyle's rundown.
Our take: the 600-to-2 pipeline
In April, more than 600 employees signed a letter warning their chief executive exactly where this road ends. The company's answer was a guardrail word — should — buried in a contract whose binding sentence says any lawful government purpose. Six months later, the named exits began. That is not a coincidence; it is a pipeline. Letter ignored, principles eroded, people leave.
Are two resignations going to change Google's defence business? No — and pretending otherwise would be dishonest. But that is the wrong question. The right one is whether Google's much-advertised internal feedback channels work, and the evidence now points the other way: a spokesperson insists there are dedicated channels to raise concerns safely while a departing engineer says, on the record, that people feel watched and cannot speak. One of those statements is a template. The other cost someone her career.
The 2018 playbook — protest, pledge, principles — produced a document the company has since quietly gutted. If history is any guide, these resignations will not move the share price, but they will be cited the next time — and there will be a next time — as the moment staff stopped believing the principles were load-bearing. The shoddy part is not the technology. It is the governance: a company selling superintelligence that cannot, by its own researchers' account, see how its products are used once the Pentagon takes delivery. That is not oversight. It is a hope and a contract.
Quick answers
Who is Sofia Liguori?
Sofia Liguori is a DeepMind researcher who resigned over Google's Pentagon AI contracts and broke a lifetime non-disparagement clause to tell the BBC why, as covered in the article above.
What is Project Nimbus?
Project Nimbus is the 2024 cloud-and-AI contract with Israel, worth $1.2B, that sits at the centre of the DeepMind resignations covered in this article.
What is a non-disparagement clause?
A non-disparagement clause is a contractual term that bars a former employee from criticising their former employer — in Liguori's case, reportedly for life, which she broke to speak to the BBC.
Who resigned from Google DeepMind? Research engineer Sofia Liguori (London) resigned in the week of 19–21 October 2026 after three years, and AI researcher Andreas Kirsch resigned earlier in October. The BBC reported several more tech workers are in the process of leaving over the same concerns.
What does Google's Pentagon contract actually allow? The contract, signed earlier in 2026, permits Google's AI models to be used for any lawful government purpose. Guardrail language says models should not support domestic mass surveillance or autonomous weapons without appropriate human oversight and control.
What did the April employee letter demand? On 27 April 2026, more than 600 Google employees wrote to CEO Sundar Pichai demanding the company reject classified military workloads outright — the only way, they argued, to guarantee Google would not become associated with lethal autonomous weapons or mass surveillance.
Sources: This article relies on the BBC's 21 October reporting (via its syndication relays the Nile Post, Ada Derana, KahawaTungu and when2meet.net), AFP reporting via The Hindu and the Straits Times, the Washington Post's reporting on the April employee letter (via the Times of India), Bloomberg Law's first report of that letter, The Intercept's 21 October interview with Sofia Liguori (via AI Weekly and Superpower Daily), and the dailyaibrief summary.
Daniel Rao is TechPulse USA's AI-policy correspondent, covering the collision between Silicon Valley and the Pentagon. He previously reported on enterprise technology in New York and London.
Comments
Post a Comment