In a nutshell
More than 560 Google employees, including senior figures from DeepMind, have signed an open letter urging CEO Sundar Pichai to refuse classified US military and intelligence contracts involving the company’s AI models. According to the Financial Times, the letter responds to reports that Google is nearing a Department of Defense agreement that would deploy Gemini in classified operations without the guardrails Anthropic had demanded: restrictions on lethal autonomous weapons and mass domestic surveillance.
The wider context is unusually charged. After Anthropic CEO Dario Amodei declined to grant the US government unrestricted access to Claude, the company was designated a supply-chain risk and federal departments were ordered to stop using its products. This is being challenged in court. OpenAI struck a more permissive deal and faced internal backlash, prompting a public apology from its CEO. Google also removed, in a 2025 update to its AI Principles, earlier language pledging not to pursue AI applications involving weapons or surveillance.
Our take
It is tempting to read this story as a clean morality tale: a coercive administration on one side, principled employees on the other.
We are not convinced, even though the asymmetry matters. State action and employee voice are not equivalent forms of power: governments can designate, exclude, and condition procurement in ways that move entire markets. Employees, on the other hand, can petition and impose reputational or retention costs on a single firm at a time. The Trump administration’s posture toward Anthropic would well be the more troubling vector, and could amount to the kind of politicization of foundational infrastructure we exist to oppose. Employee petitioning, by contrast, is a legitimate form of internal voice. The concern is not that staff speak, of course, but that corporate governance over network-essential technology may become hostage to whichever faction pushes hardest.
The deeper question is therefore procedural: what legitimate mechanisms should govern the deployment of network-essential AI? Government may fail the tests of due process and proportionality. Staff petitions, however conscientious, are not a substitute for institutional governance either. A critical infrastructure provider cannot be run by whichever external or internal constituency applies the most pressure. Until clear, accountable rules exist for AI deployment in military and intelligence contexts, such as statutory authority, transparent categories of prohibited use, independent oversight, proportionality standards, it is likely that each round of this fight will make the infrastructure layer more politicized, not less.
