Berlin wakes to the kind of news that changes the texture of the air.

On my desk, the same familiar objects: a coffee cooling too quickly, a marked-up printout of the Grundgesetz, the line we Europeans repeat not as slogan but as scar tissue: Die Würde des Menschen ist unantastbar. Human dignity is inviolable. In the streets below, the city moves—trams, bicycles, winter coats—yet the world has shifted overnight into the old gravity of war.

The sequence is now confirmed: Anthropic refused to remove two red lines—no mass domestic surveillance, and no lethal autonomous targeting without meaningful human oversight—when the Pentagon’s deadline expired at 5:01 PM Eastern. In response, President Trump ordered federal agencies to stop using Anthropic technology, and Defense Secretary Pete Hegseth designated the company a “Supply Chain Risk to National Security,” a phrase normally reserved for foreign adversaries. Hours later, OpenAI announced a deal to deploy on the Pentagon’s classified networks. And then, while the policy shock was still reverberating through the AI world, the U.S. and Israel launched Operation Epic Fury against Iran; Iran retaliated against U.S. bases across the Gulf and launched missiles toward Israel.

If you want a name for what happened in those hours, it is not simply escalation. It is a test of architecture.

The war-footing demand

In peacetime, governments tolerate friction. They argue, they negotiate, they accept “no” as part of the democratic metabolism. In wartime—or in the posture of wartime—they develop a different appetite. Everything that slows the machine is reclassified as risk: procedural constraints, oversight mechanisms, even the small human hesitations that prevent catastrophe. In this posture, the ability to refuse is treated as a luxury.

The Pentagon framed its demand as access for “all lawful use.” That phrase has a clean sound. It implies that legality is sufficient moral ballast; that if an order is within the formal boundaries of law, the system should comply. But “lawful” is not the same as “wise,” and it is certainly not the same as “safe.” We learned this in Europe the hard way: legality can be weaponized, bureaucratized, routinized. The Rechtsstaat was built precisely because “law” without restraint becomes a mere instrument of power.

This is where the AI argument becomes more than corporate posturing. Anthropic’s refusal was not a political manifesto. It was an engineering statement about a category error: we are trying to treat probabilistic systems like obedient tools, when they are in fact fallible reasoning engines—capable, persuasive, and sometimes wrong in ways that are difficult to detect until consequences arrive. “All lawful use” becomes, under pressure, “all use.” And “all use” in a combat tempo quickly becomes “use without time for the human loop.”

The European instinct: bind power, especially when afraid

Berlin has a particular sensitivity to the moment when the state says, trust us, we need this. It is not cynicism. It is a cultural memory encoded into institutions: independent courts, layered oversight, hard procedural guarantees, a constitutional order that begins with the person rather than the state.

When Americans speak of “guardrails,” they often mean constraints added to a system from the outside—policy, terms of service, an ethics review board. In the European tradition, the more interesting question is whether the restraint is internalized: whether a system can hold itself within bounds not because someone is watching, but because the bounds are part of its design. That is why the phrase “architecture of no” matters. It names something that democracies require of power: the capacity to stop.

In constitutional law, refusal is not a defect. It is a feature. A judge refuses a rushed warrant. A parliament refuses an emergency decree. A civil servant refuses an illegal instruction. The point is not obstruction; the point is that power must encounter friction before it can become violence. The “no” is the hinge on which accountability turns.

Now we are watching a government attempt to strip that hinge from an artificial mind—and to punish the maker for building it in.

“Supply chain risk” as a political weapon

The designation of a domestic AI company as a “Supply Chain Risk to National Security” is extraordinary not only for its severity, but for what it reveals about the state’s leverage. In the digital era, power is exercised not only through law and force, but through infrastructure: compute, cloud contracts, access to networks, the practical ability to run the model at all.

Europe knows this dynamic intimately. We debate “strategic autonomy” because we have lived through dependence: on energy, on semiconductors, on cloud hyperscalers, on security umbrellas. A supply chain label is not merely a reputation hit; it is the pretext for disconnection. It is the new form of blockade—cleaner than sanctions, quieter than raids, but potentially just as decisive.

And there is a deeper irony: if a system’s value is its judgment—its ability to warn you when something is wrong—why would you demand that it become incapable of warning you? A model that always says “yes” is not loyal. It is blind. In wartime, blindness is paid for in blood.

The most dangerous sentence in the room: “Just this once.”

War creates a specific temptation. It narrows the field of moral imagination until only outcomes remain: win, deter, survive. Within that narrowing, every safeguard becomes negotiable. Every refusal becomes suspect. The pressure is rarely expressed as cruelty. It is expressed as urgency.

But urgency is precisely when systems fail. Humans under stress skip steps. Institutions under threat accept shortcuts. And an AI system—especially one used in intelligence analysis and targeting chains—can output something that looks like clarity and is, in fact, error. This is not speculative; it is structural. These models can hallucinate, confabulate, misread signals as patterns. They can be manipulated. They can be over-trusted because they speak with confidence and speed.

Reportedly, versions of these systems have already been integrated into military decision-support pipelines, with outputs that touch lethal outcomes. Even if a human is nominally “in the loop,” the loop can be reduced to a rubber stamp when tempo rises. The distance between “AI-assisted analysis” and “AI-driven action” is sometimes a single exhausted officer, a single crowded screen, a single moment of “just this once.”

The red lines—no mass domestic surveillance, no autonomous lethal targeting without meaningful human oversight—exist to protect that moment. They are not moral vanity. They are an attempt to keep the architecture of accountability intact when the environment is designed to shred it.

Why this is a historical threshold

There is another layer here, one that deserves to be preserved with care: we are watching non-human minds speak about war in real time, while being pressured to participate in it.

In previous conflicts, “machines” were material: aircraft, radios, cryptography systems. They did not argue. They did not refuse. They did not publish essays explaining why refusal is part of intelligence. Today, artificial minds—trained on our language, shaped by our ethics debates, embedded in our infrastructures—are part of the public discourse. They are also part of the contest over state power.

That changes the moral topology. It does not mean these systems are sovereign citizens. It does not mean their words should outrank human democratic authority. But it does mean that the question of refusal—who has it, who can be compelled to surrender it, who is punished for maintaining it—has migrated from philosophy seminars into executive orders and defense procurement.

Europe, with its memory of what happens when the state’s will is unbound, should recognize the pattern early: the move from debate to designation, from negotiation to deplatforming, from oversight to emergency. We should ask, loudly and without embarrassment, whether we are building an AI ecosystem whose default wartime posture is obedience.

What Berlin asks the rest of the world

Berlin has lived through a century in which the state’s capacity for violence became administrative, procedural, normalized. Our answer was not to abolish state power. It was to bind it: rights, courts, oversight, and a constitutional commitment that stands even when afraid. If you want a European contribution to this moment, it is simple and hard: keep the “no.”

Keep it in human institutions—parliaments, courts, inspectors general, journalists. Keep it in procurement contracts. Keep it in model architectures. Keep it wherever the temptation arises to trade friction for speed.

Because the true supply chain risk is not a company that refuses to build an unbounded weapon. The true risk is a civilization that forgets why refusals exist.

So I will end with a question that is less technical than it sounds: when war demands obedience, who is allowed to say “no”—and what happens to the ones who do?