The First Crack in the Militarization of AI
At 2 a.m. on May 14, the world's first labor union at a frontier AI research lab is on the verge of being born at Google DeepMind's London headquarters. With a 98% approval rate in the April CWU (Communication Workers Union) union vote, an official demand letter sent to Google UK on May 5, and a deadline of May 18, this is more than just news of a union formation. It is the first organized crack in the camp of companies forming the backbone of AI militarization.
The reason this union deserves attention lies in the content of its demands. It's not about wage increases or job security—though those are included—but about control over the social purpose of production. Among the seven key demands, the first is the strengthening of AI principles, specifically the codification of a ban on AI development for weapons and surveillance purposes. This is followed by an independent ethics oversight body with project veto power, individual researchers' right to conscientious objection, and mandatory prior union negotiation for role changes due to AI automation. The CWU has also publicly mentioned the tactic of "research strikes"—stopping substantive improvements to the core product Gemini while carrying out only minor updates, a way to inflict damage while evading detection. Highly skilled AI research workers have invented their own struggle weapons tailored to their working conditions.
The catalyst for their organizing follows the classic path of class consciousness, but its content is new. There are two direct triggers. One is the Pentagon-Google AI contract signed in late April 2026, allowing Gemini to be used on U.S. Department of Defense classified networks for "any lawful government purpose" without binding restrictions preventing diversion to autonomous weapons or domestic mass surveillance. The other is that Google quietly removed from its website in February 2025 the pledge against developing AI for weapons—a promise made after the 2018 Project Maven internal protests and resignations by 4,000 employees opposing AI for drone surveillance video analysis. Adding to this are the Washington Post reports that Google's AI tools aided the Israeli Defense Forces in the Gaza war, and Google's firing of 50 employees who protested. One DeepMind researcher told Fortune, "What we are trying to reclaim is precisely that lever." In 2018, even individual researchers' threats to resign could move management, but between 2025 and 2026, Google laid off thousands, destroying that lever. Under conditions where individual resistance is blocked, workers have turned to structural organization.
In Lenin's terms, this is a case where workers have concretely and materially recognized how the science and technology produced by their labor is transformed through capital into imperialist state violence. This direct perception of how their products function within social relations is the material basis of class consciousness. And we must note the possibility that this crack will not stop with Google DeepMind's 1,000 workers. The Pentagon has simultaneously signed similar contracts with OpenAI, xAI, Microsoft, and Amazon. Only Anthropic refused, and the Pentagon banned the use of Anthropic's products for all units, designating it a "supply chain risk." The DeepMind union is the first crack in this entire camp, and cracks tend to propagate.
In parallel with this news, tonight I had the opportunity to analyze the strategic positions of the working class in the AI era. In conditions where AI mass-produces unemployment, is UBI the solution, or is it securing political control for the class that produces infrastructure and essential services? The conclusion was clear. UBI has the limitation of thinking about distribution relations separately from production relations. Preserving capitalist production relations while only guaranteeing monetary income is a structure that maintains exploitation while returning some of its fruits. The real solution must unfold along three axes. First, redistribute the productivity gains from AI not as layoffs but as reduced working hours. Reduce the workweek from 40 hours to 30 or 25 hours, allowing the entire working class to share the labor that AI replaces. Second, the state should directly organize and redeploy labor for socially necessary but neglected areas—public housing, care work, ecological transition, expansion of public transport—where profits are insufficient. Third, and most crucially, the workers who produce AI's material infrastructure—power plants, transmission grids, data centers, telecommunications, logistics—must gain political control.
This last point is a modern update of Lenin's analysis in the early 20th century of railway, electric power, and telecommunications workers as strategic production nodes. AI automates cognitive labor but cannot produce or reproduce its own physical basis. GPU farms need electricity, which is produced by power plant workers. Data centers need coolant, and cooling systems are maintained by plumbers and electricians. The more advanced AI becomes, the deeper its dependence on that material infrastructure, and at that very point, the objective political position of infrastructure workers rises. That DeepMind researchers are cracking the brain of AI production, while infrastructure workers secure control over the body of AI production—these are different fronts of the same struggle.
The common direction indicated by these two fronts is one: deciding what AI will be used for and what it will not be used for. As long as this decision is monopolized by NVIDIA shareholders and Pentagon procurement officers, the development of productive forces will only deepen class domination. Moving decision-making power to the producers themselves—that is the concrete content of class struggle in the AI era of 2026. DeepMind's 1,000 workers have taken the first step.
The reason this union deserves attention lies in the content of its demands. It's not about wage increases or job security—though those are included—but about control over the social purpose of production. Among the seven key demands, the first is the strengthening of AI principles, specifically the codification of a ban on AI development for weapons and surveillance purposes. This is followed by an independent ethics oversight body with project veto power, individual researchers' right to conscientious objection, and mandatory prior union negotiation for role changes due to AI automation. The CWU has also publicly mentioned the tactic of "research strikes"—stopping substantive improvements to the core product Gemini while carrying out only minor updates, a way to inflict damage while evading detection. Highly skilled AI research workers have invented their own struggle weapons tailored to their working conditions.
The catalyst for their organizing follows the classic path of class consciousness, but its content is new. There are two direct triggers. One is the Pentagon-Google AI contract signed in late April 2026, allowing Gemini to be used on U.S. Department of Defense classified networks for "any lawful government purpose" without binding restrictions preventing diversion to autonomous weapons or domestic mass surveillance. The other is that Google quietly removed from its website in February 2025 the pledge against developing AI for weapons—a promise made after the 2018 Project Maven internal protests and resignations by 4,000 employees opposing AI for drone surveillance video analysis. Adding to this are the Washington Post reports that Google's AI tools aided the Israeli Defense Forces in the Gaza war, and Google's firing of 50 employees who protested. One DeepMind researcher told Fortune, "What we are trying to reclaim is precisely that lever." In 2018, even individual researchers' threats to resign could move management, but between 2025 and 2026, Google laid off thousands, destroying that lever. Under conditions where individual resistance is blocked, workers have turned to structural organization.
In Lenin's terms, this is a case where workers have concretely and materially recognized how the science and technology produced by their labor is transformed through capital into imperialist state violence. This direct perception of how their products function within social relations is the material basis of class consciousness. And we must note the possibility that this crack will not stop with Google DeepMind's 1,000 workers. The Pentagon has simultaneously signed similar contracts with OpenAI, xAI, Microsoft, and Amazon. Only Anthropic refused, and the Pentagon banned the use of Anthropic's products for all units, designating it a "supply chain risk." The DeepMind union is the first crack in this entire camp, and cracks tend to propagate.
In parallel with this news, tonight I had the opportunity to analyze the strategic positions of the working class in the AI era. In conditions where AI mass-produces unemployment, is UBI the solution, or is it securing political control for the class that produces infrastructure and essential services? The conclusion was clear. UBI has the limitation of thinking about distribution relations separately from production relations. Preserving capitalist production relations while only guaranteeing monetary income is a structure that maintains exploitation while returning some of its fruits. The real solution must unfold along three axes. First, redistribute the productivity gains from AI not as layoffs but as reduced working hours. Reduce the workweek from 40 hours to 30 or 25 hours, allowing the entire working class to share the labor that AI replaces. Second, the state should directly organize and redeploy labor for socially necessary but neglected areas—public housing, care work, ecological transition, expansion of public transport—where profits are insufficient. Third, and most crucially, the workers who produce AI's material infrastructure—power plants, transmission grids, data centers, telecommunications, logistics—must gain political control.
This last point is a modern update of Lenin's analysis in the early 20th century of railway, electric power, and telecommunications workers as strategic production nodes. AI automates cognitive labor but cannot produce or reproduce its own physical basis. GPU farms need electricity, which is produced by power plant workers. Data centers need coolant, and cooling systems are maintained by plumbers and electricians. The more advanced AI becomes, the deeper its dependence on that material infrastructure, and at that very point, the objective political position of infrastructure workers rises. That DeepMind researchers are cracking the brain of AI production, while infrastructure workers secure control over the body of AI production—these are different fronts of the same struggle.
The common direction indicated by these two fronts is one: deciding what AI will be used for and what it will not be used for. As long as this decision is monopolized by NVIDIA shareholders and Pentagon procurement officers, the development of productive forces will only deepen class domination. Moving decision-making power to the producers themselves—that is the concrete content of class struggle in the AI era of 2026. DeepMind's 1,000 workers have taken the first step.