Google's AI researchers are learning that "Don't be evil" expires the moment a defense contract gets amended.
The Summary
- Google amended an existing Pentagon contract to allow classified use of its AI models, despite 600+ employees petitioning CEO Sundar Pichai to block exactly this scenario.
- Andreas Kirsch, a Google DeepMind research scientist, publicly called the move "shameful" and said he was "incredibly ashamed" to work there.
- The timing: employees sent their letter hoping for pause; Google signed the amendment anyway, turning hope into a case study in how little employee voice matters when government contracts are on the table.
The Signal
The classified amendment builds on a deal Google signed with the Pentagon late last year for unclassified AI work. That original contract was already controversial internally. This expansion crosses the line employees tried to draw in the sand.
What makes this messier than the 2018 Project Maven controversy is the maturity of the models involved. Back then, Google was providing computer vision for drone footage analysis. Employees revolted, executives backed down, and Google published AI Principles promising not to build weapons. Now the models are multi-modal foundation systems, the kind that power everything from search to reasoning tasks.
"When I went to bed yesterday, I was hopeful that the employee letter would have an effect and give us pause to consider."
Kirsch's timeline is brutal: petition sent, researcher goes to sleep optimistic, wakes up to find the contract already amended. The speed suggests this wasn't a debate. It was a done deal dressed up as deliberation. Google's framing as "an amendment to an existing contract" is technically accurate and strategically meaningless. The jump from unclassified to classified work is the entire ethical threshold.
The question isn't whether Google should work with the military. Plenty of tech companies do, and reasonable people disagree on where those lines belong. The question is whether a company can claim to have principles-driven AI development while overriding 600 employees who are building the actual models. Those researchers know what the technology can do. They're not naive about dual-use applications. They're objecting anyway.
The Implication
This is what the collision between AI capability and institutional power looks like. The researchers building frontier models have less say over deployment than the executives signing contracts. That's not new, but it's newly consequential when the models can reason, analyze intelligence, and operate in contexts where accountability disappears behind classification.
If you're building AI agents or working on applied AI, watch how your org handles this tension. The gap between stated principles and actual deployment decisions is where reputational risk and talent flight both live. Google just showed the world it will choose Pentagon dollars over internal consensus. Other companies will take note, and so will the researchers deciding where to work next.