Google DeepMind Non-Disparagement Clause Explained

Executive Summary

The employment contracts of Google DeepMind's UK staff carry a clause barring them from criticizing the company for the rest of their lives, and the bar does not lift on the day the job ends. The Washington Post reviewed the contracts, interviewed two current employees and published the story on November 3. Both asked not to be named for fear of retaliation, and both said what the clause blocks is talking with the public about what AI will do to the world.

The clause came to light because of one person, a research engineer in London who resigned in October. She decided to break it and spoke, the colleagues who stayed pulled up their own contracts and found the same wording, and dozens of them gathered in an internal message thread. Google answered that it will not enforce the clause against anyone who has already left, and said it had removed the post-termination part from the terms it gives new hires. The staff asked for the removal of the clause from the contracts that already exist, and that request has not been granted.

Sections 1 through 3 follow what has been reported and what UK law actually protects. Section 4, where the case moves over to the side that receives AI safety documents, is this article's reading rather than anything the reporting says.

Key Numbers

The first two cells mark how far the contract that came to light reaches and how far Google has gone. The last two show what people put on the line the previous time the same request was made, in 2024.

Sources: Washington Post (2026-11-03), Right to Warn letter (2024-06-04), CNBC (2024-05-24)

For life

How long the clause runs

It does not end on the last day of employment. That is how it reads in the contracts of UK-based staff

Zero

Clauses cut from existing contracts

Google changed only the terms for recent new hires. The contracts of current staff and of people who have left stand as they were

6 of 13

Signatories who withheld their names

Of the people who signed the 2024 Right to Warn letter, this many kept their names off it, fearing the company would retaliate

About $2M

What refusing to sign put at risk

The equity a departing OpenAI researcher staked in 2024 by refusing to sign. The company then released every former employee

1

One Person Broke the Clause and the Rest Opened Their Contracts

Sofia Liguori joined Google DeepMind in 2023 and worked as a research engineer in London. Her assignment was AI for pushing electricity through the grid more efficiently, a seat far from anything military. As she came to learn that the company was tilting toward putting AI into military operations, she said, working at Google began to feel shameful. She resigned in October, three years in, and the AI researcher Andreas Kirsch left the company that same month for similar reasons.

In an interview with the BBC on October 21, Liguori opened by stating that her own contract forbids exactly that kind of speech. "In principle I'm not allowed to criticise Google for the duration of my life." She judged it was something that had to be known, and added that stating facts is not criticism. On the mood inside the company she said, "people are scared of retaliation and putting their career at risk. You feel you are being watched, you can't speak."

What the same article carried was not one person's situation. Employees at Google told the BBC that staff are afraid to speak publicly about AI safety and about how AI is used in warfare, because their contracts of employment forbid them from criticizing the company in public. Andreas Kirsch, by then out of the company, gave his own reasons in that article. Google's governance and oversight, he said, "are not sufficient or meaningful enough to deploy AI safely and responsibly in military and classified settings." Google, he added, "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."

After the interview ran, the staff who remained opened their own contracts and confirmed the same wording. Dozens gathered in an internal message thread to denounce the clause, and some reacted with the George Orwell 1984 emoji. That argument reached the outside on November 3, in the Washington Post. The article rested on a reading of the contract itself and on interviews with two current employees, and it was the first time the specifics of the clause were made public. Until then, what was known was a single sentence from Liguori.

The words of the two current employees in that report show where the demand to drop the clause comes from. One of them said, "This technology is just really consequential and it's also changing really fast. The public wants to know and deserves to know." The other said, "The people who work on this technology have a role to play in shaping where it's going. It's not just another technology." Both spoke on condition of anonymity to avoid retaliation.

2

What Google Offered and What It Did Not

The company's first answer came on October 30, in the internal thread. An executive stepped into the argument directly and said the clause would not be enforced against anyone who had left. The official position followed. Google had updated the terms for new hires, taking out the part of the non-disparagement clause that applies after termination; it had always encouraged open discussion of safety and responsibility; and internal routes exist for carrying concerns to senior leadership.

Staff answered that this was not enough. Their demand was that the clause be struck from existing contracts too, and there is a practical reason behind it. What the company offered is a statement that it will not enforce. What stays in the contract is the wording. When a former researcher wants to say what they saw at a conference or a hearing, the thing they can lean on is not their own contract but something the company said once. If the person who said it moves on, or if company policy shifts, the contract decides what is left.

How much that difference matters was settled once already, two and a half years ago, in the same industry. In May 2024 OpenAI released every departing employee from open-ended non-disparagement obligations in an internal memo. It cut the clause from its standard exit paperwork, stated in writing that it would not enforce obligations already entered into, and added a sentence saying that vested equity had not been canceled and would not be. The trigger was the refusal of Daniel Kokotajlo, a researcher who left that year, to sign. The equity his refusal could have cost him came to about $2 million.

The request in the two cases is the same and the shape of the answer is different. What OpenAI produced was a document withdrawing an obligation from contracts that already existed. What Google produced is an announcement that it will fix the contracts of people who join from now on, together with a spoken assurance that the clause will not be applied to the people who are already out. The contracts of the people still at the company fall under neither.

The demand did not come only from inside the company. Nathan Calvin, head of legal at the AI policy nonprofit Encode, wrote that it is "truly appalling" for Google to make AI researchers sign non-disparagement clauses "that last in perpetuity", and asked the company to discontinue the practice immediately and release the former employees who have already signed from those obligations. How many former employees that is, though, nobody has counted. The Washington Post did not report the number of current and former staff bound by the clause, and the clause has never been tested at a UK employment tribunal.

3

The Scope of UK Whistleblower Law, and the Concerns Outside It

The legal force of the clause is worth settling first. The UK has a law protecting public interest whistleblowers, and one of the provisions the Public Interest Disclosure Act of 1998 inserted into employment law treats a contractual term that seeks to stop a protected disclosure as void to that extent. It makes no difference whether the term was agreed during employment or on the way out. Few people therefore expect this clause to hold up in court as written, and the employee the Washington Post quoted was speaking from that same premise. "If it's not enforceable then the reason is to basically intimidate people."

The trouble is the scope of a protected disclosure. What the law protects is a criminal offense, a breach of a legal obligation, a miscarriage of justice, a danger to the health or safety of a person, damage to the environment, and the concealment of any of these. The protection works when you can name which item applies. Much of what AI researchers want to tell the outside world does not sit comfortably on that list. Take how the numbers from an evaluation were selected and which of them were published, at what point and how far a safety check was run, or how a concern raised inside the company fed into a release decision. No law has yet set any of these down as a duty.

That gap had already been put into sentences in a letter. On June 4, 2024, thirteen current and former employees of OpenAI and Google DeepMind published the Right to Warn letter. Its argument was that ordinary whistleblower protections are aimed at unlawful conduct, while many of the risks the signatories worried about are not yet regulated. The first of the four principles they asked for was that companies neither enter into nor enforce any agreement forbidding disparagement or criticism of the company for risk-related concerns. Of the thirteen signatories, seven gave their names, and the remaining six stayed anonymous because they feared retaliation. The seven who gave their names include current and former researchers at Google DeepMind.

So this is not a new demand. The first principle that people put their names to in 2024 was still sitting inside the same company's contracts in 2026, and what it took to confirm that from the outside was one person breaking the clause and one reporter reading the contract. Whether the clause would hold up in court is the question after that. Regardless of enforcement, the person who has read the contract keeps quiet, and the further a story sits outside the protected list, the more surely that happens.

4

The Vendor Writes the Model Card

From here the article leaves the reporting behind. Think about the documents in your hands when you judge which model to adopt, and almost every one of them was written by the company that built the model. The model card is one. So is the system card, and so are safety evaluation reports and frontier safety policy documents. Which evaluations were run, which results went in, and what was left out are all settled by the same side. This is not to say the documents are poor; it means the author and the verifier sit in the same place.

The second route, the external evaluation body, is in a similar position. When access to the model begins, how far inside it they get to look, and whether results may be published are mostly fixed by an agreement with the company. It is a route that opens as far as the company opens it, so whatever the company decided not to show stays invisible there too. What is left is the third route, the testimony of the people who built the model. That is exactly the place this reporting pointed at.

▲ Original Pebblous diagram | Grounds: the Washington Post report and the Right to Warn letter

This structure is not peculiar to Google. Seen from the side that buys models, it means the people able to answer back to a vendor's safety claims are the people who built the model, and their hands may be tied. Here is the point where a line in a contract stops being that company's reputation problem and becomes material for the risk assessment of whoever adopts its models. Three questions are worth asking when choosing a vendor.

Are the people who wrote the safety documents we received and the people who verified them in the same organization? If they differ, what could the verifying side see, and from when?
Can current and former employees of that company speak publicly about risk? Does a non-disparagement clause that survives their departure remain in their contracts?
Is the internal channel the only route for raising a risk? When that channel closes, does the contract leave a way out?

Given the state of the industry, a yes to all three is hard to come by. Even so, a record that the question was asked becomes grounds for fixing one line in the next contract. The change at OpenAI in 2024 also came of being asked from the outside.

Editor's Note

What Pebblous looks at first when diagnosing data quality is not the number but who wrote it down and under what conditions. The same question applies to AI safety material. When an evaluation result arrives from a vendor, recording which organization wrote it and whether a route is open for answering it back from the outside adds one more line of grounds to the decision to adopt a model.

Thanks for reading this far. The original report is at the Washington Post, and summaries ran in the English editions of Seoul Economic Daily and SBS. The longer interview with Sofia Liguori is at Democracy Now. If your team already has the speaking freedom of current and former employees in its vendor assessment, we would like to hear the wording you use to ask.

R

References