Alex Turner, a research scientist who spent more than two years on Google DeepMind’s alignment team, has published a detailed account of why he quit the company this spring — a classified Pentagon contract for Gemini that he says leaves the military’s use of the model without real limits.
What the contract allows
Google signed the agreement with the U.S. Department of Defense on April 27, according to Turner, granting the Pentagon use of Gemini for “any lawful government purpose,” including military applications. The contract states the system “should not be used for domestic mass surveillance or autonomous weapons… without appropriate human oversight and control” — but Turner says that language isn’t binding, and the deal separately confirms Google has no right to veto a lawful government operational decision.
A rejected proposal
Turner says he spent months trying to change the outcome from inside Google. He wrote a 25-page framework proposing enforceable safeguards — including mandatory human control over any targeting decisions and a seven-person review body reporting to the company’s chief scientist — and sent it to DeepMind chief executive Demis Hassabis and chief scientist Jeff Dean. He also helped organize an internal petition that drew roughly 600 signatures from employees urging the company not to sign. None of it changed the contract’s terms.
“When Google signed, I just couldn’t do any more work,” Turner wrote. “My brain said ’no.’” He left soon after and says he has since turned down interest from rival AI labs.
A Google spokesperson told Business Insider the company had been receptive to Turner’s proposals and remains committed to the view that AI should not be used for domestic mass surveillance or autonomous weapons without human oversight.
Part of a wider pattern
The resignation follows Google’s decision, in early 2025, to drop a pledge against building AI for weapons from its published AI principles, and lands amid mounting employee ethics concerns over military AI contracts across the industry — an area DeepMind’s own CEO has separately proposed putting under a dedicated safety watchdog. Turner’s account also describes internal alarm earlier this year over a similar Pentagon push aimed at Anthropic, Google’s closest AI rival.