What Google's Employees Were Promised, And What They Got

On 28 April, Google expanded its existing $200 million Pentagon contract to allow Gemini to be used on classified systems for "any lawful governmental purpose."¹ The phrase is doing a lot of work. It is the same language OpenAI and xAI accepted in their own defence agreements, and it is, in practical terms, a blank cheque written against the model's future uses.² Within days, more than 600 Google employees, a substantial bloc of them from DeepMind, signed a letter to Sundar Pichai protesting the decision. Senior researchers told reporters that leadership had privately assured them, in conversations during the recruitment and retention battles of the last two years, that Google would not sign exactly this kind of deal.³

I want to be careful about what the story here is and what it isn't.

It isn't that Google has done something unprecedented. It hasn't. The defence-AI integration has been visible for at least three years, and the question was never whether the major labs would be on classified systems but when and on what terms. It isn't, either, that "any lawful governmental purpose" is a shocking phrase on its face. Lawful is lawful. The contract is public. There is no smoking gun in the text.

The story is narrower and, I think, more important. It is about what a particular group of workers were told, what they did on the strength of being told it, and what happened when the strategic situation changed.

The DeepMind researchers who pushed back hardest are, by most accounts, people who chose Google specifically because of commitments made after the 2018 Maven protests, commitments Google itself codified in its AI Principles, which originally barred work on weapons and on technologies "whose principal purpose or implementation" would cause harm.⁴ Those principles were quietly revised in early 2025 to remove the explicit weapons carve-out.⁵ The April expansion is the operational consequence of that revision. The researchers who are angry now are angry because they took the original commitments as part of the compensation package, not in dollars, but in the moral architecture of the work. They are discovering that the architecture was load-bearing only as long as it was cheap.

This is a pattern worth naming, because it is going to keep happening.

The labs are in a war for capital and for government contracts that they cannot afford to lose to one another. In that environment, prior ethical commitments function as constraints on competitive strategy, and constraints on competitive strategy get revised. The revisions are presented as updates, as more sophisticated readings of the same underlying values, suited to a more complicated geopolitical moment. Sometimes that is even sincere. But the workers who joined on the basis of the older readings are not wrong to feel that something has been taken from them without their consent.

And consent is the right frame. When Google reassured senior researchers that it would not pursue classified defence integration, those reassurances had material effects: people stayed, people joined, people turned down offers from Anthropic and from US national-security-focused startups. The reassurances were, functionally, terms of employment, unwritten, but relied upon. Their unilateral revision is a kind of taking. The fact that it is legal does not make it neutral.

The wider question, the one I keep coming back to, is who actually gets to decide what these systems are for. The contract language, "any lawful governmental purpose", is a deliberate transfer of that decision from the lab to the customer. Google is not, under this contract, deciding case by case which classified uses of Gemini it endorses. It has pre-endorsed the set. The engineers who built the model, including those who built the safety stack on top of it, have no further say. Neither, of course, do the people on the receiving end of whatever the lawful purposes turn out to be.

I don't think this is the worst thing happening in AI right now. I think it is one of the more legible examples of a thing that is happening across the industry: the slow conversion of ethical commitments into marketing assets, retained when convenient and retired when not. The 600-plus signatories of the letter are doing the only thing left to them, which is to put on the public record that they were told something different. Whether that record matters depends on whether anyone, regulators, future hires, the press, treats it as evidence of a pattern rather than a one-off grievance.

I think it is evidence of a pattern. The Maven protest worked, briefly, because Google's leadership decided the cost of losing the researchers exceeded the value of the contract. That calculation has now flipped. The researchers are still here; the contract is bigger; the principles have been amended to fit. What changed was not the ethics. What changed was the price.

Footnotes

¶