When AI Says No to the Pentagon: Anthropic, the Department of War and Claude's Constitution

On February 26, 2026, Dario Amodei — CEO and co-founder of Anthropic — published a statement destined to make waves. The company behind Claude, one of the most advanced language models in the world, announced that it had refused certain requests from the United States Department of War. This was not a blanket refusal to work with the military — Anthropic was the first frontier AI company to deploy its models on classified government networks — but a denial on two specific cases: mass surveillance of American citizens and fully autonomous weapons without human oversight.

The Pentagon's response was swift. The Department of War threatened to remove Anthropic from its systems, designate it as a "supply chain risk" — a label normally reserved for foreign adversaries — and invoke the Defense Production Act to force the company to remove its limits. Amodei called these threats "inherently contradictory": you cannot simultaneously be a national security risk and an essential supplier for critical military operations.

A position rooted in values, not just contracts

Beyond geopolitical and commercial dynamics, there is a deeper question worth asking: where does this line that Anthropic is unwilling to cross concretely come from?

A plausible answer might be found in the so-called Claude's Constitution, the document — published in its most recent version in January 2026 — that defines the model's fundamental values, priority hierarchy and absolute constraints. It is a text of about 80 pages, unusual in the tech landscape: not a simple list of rules, but an attempt to explain why certain behaviors are desirable, so the model can generalize ethical reasoning to new situations.

The Constitution establishes a clear hierarchy: safety and human oversight come before ethics, which comes before operational guidelines, which comes before commercial utility. Claude, it reads, should preserve "functioning societal structures, democratic institutions and human oversight mechanisms" and resist "problematic concentrations of power." These are principles that seem written with situations like this in mind.

Domestic mass surveillance, according to Amodei, would allow the automatic assembly at enormous scale of dispersed data on citizens' private lives — effectively circumventing existing legal protections that AI technology has now rendered inadequate. Fully autonomous weapons, on the other hand, would eliminate that human oversight the Constitution considers non-negotiable in the current phase of AI development. The argument is not ideological but technical and pragmatic: current models are not reliable enough to make lethal decisions without a human operator in the loop.

An (almost) perfect consistency — with some tension

Anthropic's refusal thus appears as a concrete application of the Constitution's principles: a case where the values declared in the training document produced real consequences, even at the cost of losing significant government contracts. It would be an admirable story of consistency, were it not for a side note that complicates the picture.

When asked whether the Constitution also applies to models deployed with the military, an Anthropic spokesperson replied that these "would not necessarily be trained on the same Constitution." This raises a legitimate question: to what extent do the values encoded in the document serve as absolute constraints, and to what extent are they flexible principles depending on the client?

For now, Anthropic seems to have held the line on the two most sensitive points. But the Pentagon controversy illuminates a structural tension that will accompany the entire AI industry in the coming years: the tension between the need for funding and contracts to sustain research, and the credibility of those who position themselves as guarantors of responsible technological development. Claude's Constitution could be, in this sense, much more than a training document: it could become the test bench on which to measure how much those promises are truly worth.