More than 580 Google employees, including senior DeepMind researchers and over 20 directors, senior directors, and vice presidents, signed an open letter to CEO Sundar Pichai on Monday demanding the company refuse to do classified AI work for the Pentagon.

Two months ago, the Department of Defense dropped Anthropic from its approved AI vendor list because Anthropic asked for guardrails against domestic mass surveillance and autonomous weapons. OpenAI signed its own Pentagon deal hours later with three voluntary red lines that mostly mirror what Anthropic was punished for asking. Google now sits in the middle of that pattern, negotiating classified access for Gemini under terms reportedly described internally as 'all lawful uses.'

The letter, first reported by Bloomberg, is unusually direct. 'Currently, the only way to guarantee that Google does not become associated with such harms is to reject any classified workloads,' it reads. 'Otherwise, such uses may occur without our knowledge or the power to stop them.'

The Air-Gap Problem

Here is what makes the classified ask different from a normal cloud contract. Classified Pentagon networks are air-gapped. They do not touch the public internet. Once Google deploys Gemini onto a classified system, Google has no way to monitor what queries are being run, what outputs are being acted on, or whether the model is being used inside an autonomous targeting loop. The only guardrail is the Defense Department's promise to use it appropriately.

This is the same structural problem Anthropic raised before getting kicked off the vendor list. The Anthropic position was that you cannot enforce a usage policy on a system you cannot see. The DoD response, basically, was: then we do not want your model.

Google appears to have studied that outcome and decided to go the opposite direction. The signatories argue this is a deliberate choice the company spent years engineering. In 2018, roughly 4,000 Google workers signed a petition against Project Maven, the Pentagon AI program for analyzing drone footage. At least twelve resigned. Google backed off, published AI principles forbidding weapons work, and let the Maven contract lapse.

The Quiet Walk-Back

Then Google quietly walked all of it back. The 'no weapons' language was removed from the AI principles in early 2025. Google won a share of the $9 billion Joint Warfighting Cloud Capability contract. Gemini has been deployed to about 3 million Pentagon personnel for unclassified work. The classified deal now under negotiation, according to The Information, would put Gemini onto the networks where AI use cannot be audited from outside.

That is the seven-year arc the new letter is pushing back on. Win the principles fight in 2018. Erode the principles for seven years. Cash the contract in 2026.

What Pichai Actually Decides

This is not a hard call to predict. Pichai has been clear that Google is 'a national champion' and intends to compete for federal AI work. The CFO has been clear that government and defense are growth lines investors expect Google to pursue. The board has signed off on the trajectory. A 580-person letter, even with VPs on it, is not the leverage 4,000 signatures plus a dozen resignations was in 2018.

What changed is the context around the letter. Anthropic just lost the contract. OpenAI just signed a different one. Google is the test case for whether a hyperscaler can take the Pentagon's money and keep its workforce.

If Pichai signs the classified deal, the public version of his answer will include phrases like 'rigorous review process,' 'consistent with our AI principles,' and 'robust ethical framework.' The internal answer will be that the AI principles got rewritten in 2025 specifically so this deal could be signed in 2026.

The DeepMind Wrinkle

The interesting signal is the DeepMind names on the letter. DeepMind has historically operated as a semi-independent research culture, with founder Demis Hassabis on record opposing weapons applications of AI. Senior DeepMind researchers signing an open letter against their own CEO is a governance signal Pichai cannot ignore even if he proceeds.

This is also where the talent-flight risk becomes real. The Big Tech-to-startup brain drain is already a 2026 story. Ex-DeepMind VP Koray Kavukcuoglu raised $1B for AMI Labs. Ex-DeepMind RL lead David Silver just raised $1.1B for Ineffable Intelligence. Both rounds were announced in the last month. If 50 senior DeepMind researchers decide the classified Pentagon contract is the line, the next press cycle is who poached them.

The Bigger Pattern

Three of the four largest American AI labs have now had their workforce force a public conversation about military AI. Anthropic took the principled stand, lost the contract, and is now publicly framed as the safety company. OpenAI took the contract, accepted the red lines voluntarily, and is now publicly framed as the pragmatic patriotic company. Google is being asked, by 580 of its own employees, to pick a lane.

Whatever Pichai decides, that decision becomes the new floor for what the rest of the industry can do. If Google ships classified Gemini, Anthropic's principled refusal looks naive. If Google walks away, OpenAI's 'all lawful uses' position becomes the outlier. The big labs do not get to be both the moral authority on AI and the prime contractor on classified military AI. That window closed two months ago.

Source: Bloomberg (original letter scoop), The Washington Post (signatory count, DeepMind), Business Insider (Information report on Gemini classified terms), CBS News (employee letter coverage), The Hill (Anthropic context), TheNextWeb (Project Maven historical context, AI principles walk-back).