Some of the people who build Google’s most advanced AI models have signed employment contracts that bar them from criticizing the company for the rest of their lives, and a group of them now wants out. According to a Washington Post report published on November 3, 2026, staff at Google DeepMind in the UK have been pressing executives to strip non-disparagement language from their existing contracts. Their argument is that people who see AI risks up close have to be able to tell the public about them.
Google’s response so far has been partial. An executive told staff in an internal discussion that the company would not enforce the clause against people who have already left. A spokesperson told the Post that Google has “always encouraged an open dialogue about safety and responsibility” and has “updated our terms for new hires, which included removing the post-termination aspect of non-disparagement clauses.” The employees say that is not enough, because the people currently doing the work are still bound by the old wording.
The account rests on contract language the Post reviewed and interviews with two current employees who spoke anonymously. Google has not disputed that the clauses exist, only, implicitly, whether they matter. Legal teams at any AI developer should be asking the same question about their own paperwork.
How the dispute started
The trigger was a public resignation. In October, former DeepMind engineer Sofia Liguori told the BBC that she had left over Google’s military AI contracts and that her contract contained a non-disparagement clause she had decided not to be bound by. Current UK staff then checked their own contracts and, according to the Post, found similar lifetime language. One employee described the technology as “really consequential” and said the public “deserves to know,” adding that disclosures by insiders are often the only way people learn about harms.
The industry has been here before. In 2024, OpenAI released former staff from exit agreements that tied non-disparagement terms to vested equity, and current and former OpenAI and Google DeepMind employees then published an open letter calling for a “right to warn” about advanced AI. The DeepMind dispute is a narrower version of the same fight, about wording in ordinary UK employment contracts.
Allegation, assurance and what is actually settled
It helps to separate the pieces. What has been reported, and not denied, is that some existing DeepMind contracts in the UK include non-disparagement obligations that continue after employment ends. What Google has said is that it does not enforce the post-termination part and has removed it from new-hire terms. What remains unsettled is the position of current employees on older contracts. An assurance given in an internal thread is not a contract amendment, and Google’s public statement does not say what would happen if a current employee publicly criticized the company’s safety practices.
One employee put the objection bluntly: if the clause is not enforceable, “the reason is to basically intimidate people.” That is the core of the complaint. Legal teams often treat broad restrictive language as harmless because the law voids it in the cases that matter. Employees reading the contract do not know that, and the chilling effect works whether or not a court would ever uphold the clause.
What the law already says about gag clauses and AI risk
Several legal regimes already limit what a contract like this can do. None of them gives employees an unrestricted right to speak to the press, and the details matter for anyone drafting or reviewing these agreements.
United Kingdom
The affected staff are in Britain, so UK law comes first. Under section 43J of the Employment Rights Act 1996, added by the Public Interest Disclosure Act 1998, any provision in an agreement is void so far as it purports to stop a worker from making a protected disclosure. The protection is real but structured. A disclosure must concern specified kinds of wrongdoing, such as a criminal offence, a breach of a legal obligation or a danger to health and safety, and the worker must reasonably believe it is in the public interest. Disclosures to an employer or a prescribed regulator are the easiest to protect. Going to the media is protected only under stricter conditions, for example where internal routes have failed or the worker reasonably fears retaliation. A general warning that a technology is moving too fast may not fit any of those categories at all.
California
California’s Transparency in Frontier Artificial Intelligence Act, SB 53, added AI-specific whistleblower protections to the Labor Code effective January 1, 2026. Section 1107.1 bars a frontier developer from making or enforcing any contract that prevents a “covered employee” from disclosing information to the Attorney General, a federal authority, a manager, or another employee with authority to investigate, if the employee has reasonable cause to believe the developer’s activities pose a specific and substantial danger to public health or safety from a catastrophic risk, or that the developer has violated the act. Large frontier developers must also run an anonymous internal reporting process with monthly status updates to the person who reported.
The limits are significant. “Covered employee” means someone responsible for assessing, managing or addressing the risk of critical safety incidents, not every researcher. “Catastrophic risk” is defined narrowly, around incidents that could kill or seriously injure more than 50 people or cause more than $1 billion in damage. The protected recipients are regulators and internal channels, not journalists. And the statute governs conduct tied to California, so its reach over UK-based employees on UK contracts is doubtful. California readers may recall that the state’s privacy regulator has separately backed a proposal for privacy whistleblower protections, which would extend similar ideas to data protection violations.
European Union
Under Article 87 of the EU AI Act, the EU Whistleblower Directive (Directive (EU) 2019/1937) applies to reports of AI Act infringements and to the people who make them. That article applies from August 2, 2026. The Directive requires reporting channels, confidentiality and protection against retaliation, and it generally makes contractual waivers of those rights ineffective. The European Commission’s AI Office has also opened a whistleblower tool for reporting suspected breaches. Again, the protection attaches to reports about infringements of the law.
Why this matters beyond Google
The gap the DeepMind staff are pointing at sits between what the law protects and what they want to say. Whistleblower statutes protect reports of unlawful conduct or specific dangers, made through defined channels. Many safety concerns about frontier AI do not look like that. They are judgments about pace, testing depth, deployment choices or how a lab weighs commercial pressure against evaluation results. Those are exactly the concerns a lifetime non-disparagement clause is most likely to silence, and exactly the ones statutory protections are least likely to cover.
For privacy and AI governance teams, that gap is a governance problem, not just an HR one. If the only route for an uncomfortable safety finding is a public warning that breaches a contract, the internal process has already failed. Courts and regulators are increasingly treating AI safety duties as something companies can be held to after the fact, and a record showing that staff were discouraged from raising concerns is the kind of evidence that ages badly in litigation.
What AI developers and deployers should check now
Companies that build or fine-tune AI models, and those that buy them, can test their own position with a short review:
- Inventory restrictive terms. Pull every template that contains non-disparagement, confidentiality or non-disclosure language: offer letters, employment contracts, equity plans, severance and settlement agreements, contractor terms. Note which have no end date.
- Add explicit carve-outs. Each clause should state plainly that nothing in it limits disclosures to regulators, law enforcement or other channels protected by law, including those under ERA section 43J, California Labor Code sections 1102.5 and 1107.1, and the EU Whistleblower Directive where it applies. Silence on this point is what creates fear.
- Fix existing contracts, not just new ones. A policy statement that a clause will not be enforced is weaker than a signed amendment or a written waiver sent to every affected person, current and former.
- Test the internal channel. Check whether there is an anonymous route for safety concerns, who receives reports, how quickly reporters get updates, and whether reports reach the board. A hotline that exists on paper is not proof it works.
- Connect safety reporting to governance records. Make sure AI risk reports feed the same register used for model evaluations and incident reviews, so concerns cannot sit in an HR file disconnected from the people approving releases.
Deployers have a stake too. If a model vendor’s staff are contractually discouraged from flagging problems, the buyer learns about those problems later, usually after an incident. Vendor due diligence questionnaires can reasonably ask whether a provider’s employee agreements carve out protected disclosures and whether it operates a functioning safety reporting process. That is a fair question for privacy teams involved in AI governance, which often already own vendor assessments.
Frequently asked questions
Are lifetime non-disparagement clauses legal in the UK?
They are not automatically unlawful, but they cannot stop a worker from making a protected disclosure. Section 43J of the Employment Rights Act 1996 voids any contractual term that tries to. A broad clause can still be enforceable against ordinary criticism that falls outside whistleblowing protection, which is why employees find it intimidating.
Does California’s SB 53 protect DeepMind researchers?
Probably not in this case. SB 53’s whistleblower provisions cover “covered employees” of frontier developers who report catastrophic risks or violations of the act to authorities or internal channels. Its application to UK-based staff on UK contracts is doubtful, and it does not protect disclosures to the media.
Has Google removed the clauses?
For new hires, Google says it has removed the post-termination part of the non-disparagement clause. For former employees, an executive said the company will not enforce it. For current employees on older contracts, the reported position is that the wording remains, which is what staff want changed.
Does the EU AI Act protect AI whistleblowers?
Yes, for reports of AI Act infringements. Article 87 applies the EU Whistleblower Directive to such reports from August 2, 2026, including confidentiality and anti-retaliation protection.
The clause nobody plans to use
Google’s message to its staff amounts to this: the clause is there, but it will not be used against you once you leave. The employees’ reply is that a promise not to enforce is not the same as removing the words. If Google amends existing contracts, it will have accepted that the people building frontier AI should be able to talk about its risks without asking permission. If it does not, the next DeepMind researcher with a serious concern will have to decide, as Liguori did, whether the contract is worth the risk of speaking up, and every AI lab with similar paperwork should expect the same question from its own staff.
