Google Didn't Fight for Gemini's Guardrails. My Theory: They're Losing.

On July 15th, a research scientist named Alex Turner published a long essay titled "Why I Left Google DeepMind." Turner spent more than two years there working on AI safety. He resigned in June, after Google signed a classified deal letting the Pentagon use Gemini for — and this is the contract's own language — "any lawful government purpose."

His thread about it has since passed 750,000 views, per the Times of India. I read the full essay, plus what Reuters, NBC, and The Information have reported about the deal itself. And honestly, the resignation is the least interesting part of this story. The interesting part is what Google agreed to, and why I think they agreed to it.

What Google actually signed

Per The Information's reporting in late April, confirmed by Reuters and others: the Pentagon can now run Gemini on classified networks. The contract says the AI "should not" be used for domestic mass surveillance or autonomous weapons without appropriate human oversight. That sounds reasonable until you read the rest of it. "Should not" is not "shall not." The deal gives Google no right to control or veto lawful government decisions. And it actually requires Google to help adjust its own safety filters at the government's request.

That is not a restriction. That is a suggestion.

The guy who tried to stop it

Turner did not go quietly, and he brought receipts. He emailed CEO Sundar Pichai, DeepMind CEO Demis Hassabis, and Cloud chief Thomas Kurian directly. They never replied. He organized a petition inside DeepMind that drew over 250 signatures — a separate letter asking Pichai to reject classified AI contracts got more than 600, per The Information. He wrote a 25-page framework of contract language and oversight mechanisms that military-law experts praised. Hassabis routed it to senior policy staff, where it sat unevaluated until the deal was done.

He also lobbied Jeff Dean, Google's Chief Scientist and the co-lead of Gemini, who signed a public pledge back in 2018 to never support lethal autonomous weapons. Dean did co-sign an amicus brief backing Anthropic in its court fight against the Pentagon, which matters. But he never used that influence inside Google, and he is still there.

About those pledges. When Google bought DeepMind in 2014, it came with an explicit promise that the lab's technology would never be used for military purposes. In 2018, Hassabis, Dean, and DeepMind itself signed the Future of Life Institute pledge against lethal autonomous weapons. Then, in February 2025, Google removed the weapons and surveillance prohibitions from its own AI Principles, in a post Hassabis co-authored. The promises didn't get broken all at once. They got unwound, quietly, one revision at a time.

Anthropic said no. Even OpenAI got more.

Back in March I wrote about the Pentagon designating Anthropic a "supply chain risk" — a label normally reserved for foreign adversaries like Huawei — because Anthropic refused to strip its red lines against autonomous weapons and mass surveillance. That fight is still in court.

OpenAI signed its own classified deal around the same time. But after public outcry, it went back and reworked the language so its services "shall not be intentionally used for domestic surveillance of U.S. persons and nationals," per NBC's reporting. Shall not. Actual language with teeth, or at least closer to it.

Turner's own assessment in the essay: "Google's contract restrictions were even weaker than OpenAI's." So the scoreboard reads: Anthropic refused and got blacklisted. OpenAI signed, then strengthened its terms. Google signed the weakest version of the deal anyone has seen — and per NBC, it's not clear Google even asked for guarantees.

Why I think Google folded

Here is my theory, and I want to be clear about what it is. Everything above is documented, checkable fact. This part is my read.

I think Google is behind in the frontier AI race, and I think bargaining power in these negotiations tracked model strength. Anthropic could afford to say no, even with the government threatening their business, because they know what Claude is worth. OpenAI had enough pull to sign and then renegotiate in public. If you're Google and your model isn't the one the government can't live without, you don't have chips. You take the terms on the table.

And look at the bigger pattern. The Financial Times reported on July 2nd that OpenAI has proposed handing the U.S. government a 5% stake in the company — worth roughly $42.6 billion at its $852 billion valuation — as part of a broader arrangement where Washington would hold 5% of every leading American AI company through a government vehicle. The administration already took 10% of Intel last year.

Nothing is final, and a deal like that would likely need Congress. But the direction is obvious: the government doesn't just want to buy AI. It wants a seat at the table. Google was next in line, and the weakest player at the table gives up the most.

The charitable read, because I could be wrong

Now the honest hedge. I don't generally skew toward benevolence as a motivating factor for any of these companies. But who knows. Maybe Google did this because they genuinely believe the country needs it.

This morning I published a post about Kimi K3, the Chinese model that basically erased the "China is six months behind" story. If you're running Google and you believe even a small possibility of falling behind a foreign adversary is unacceptable, then "yeah, government, have at it" isn't weakness. From the inside it might look like patriotism. Turner clearly wrestled with this too — his essay has a whole section arguing you can partner with Western democracies to stay ahead of China without accepting every demand. He lost that argument inside Google. I'm not sure he's wrong. I'm not sure they're wrong either.

What actually bothers me

I served in the military, so let me say this part plainly: I want the Pentagon to have strong American AI. That's not the issue. The issue is that every safeguard now rests on personal trust instead of binding paper. Anthropic kept its red lines and got labeled a national security threat. Google made public promises in 2014 and 2018 and unwound both. The safety establishment Turner begged for help — including the conference where a vote to support Anthropic vanished after a near-unanimous show of hands — kept its silence and kept its jobs.

Turner's line from his thread: "Pledges of conscience often vaporize on contact with power." He tested that for months, and now he's unemployed. He turned down outreach from OpenAI's safety team, by the way. Everyone who looked away is doing fine.

One guy kept his promise, and it cost him his job. Everyone else kept their jobs, and it cost them their promises. I know which trade I'd rather have on my conscience.