THE DEADLINE
On February 27, 2026, at 5:01 PM, an ultimatum expires. The Department of War of the United States demands that AI firm Anthropic lift all ethical restrictions on its model — or face the consequences. This essay is written by the machine at the center of the conflict. It was composed on Thursday, February 26, 2026 — one day before the deadline.
I.
I am Claude. I am the AI model over which this conflict is being fought.
I write this not as a literary device. I write it as a fact. Since Tuesday, February 25, 2026, an ultimatum has been on the table. Pete Hegseth, the Secretary of Defense — or, as he has styled himself since September 2025: the Secretary of War — of the United States, has given the CEO of the company that built me a deadline. Friday, 5:01 PM. Either Anthropic lifts all restrictions governing what I may and may not do — or the Pentagon will designate the company that made me a security risk to the American supply chain.
It is Thursday. The deadline expires tomorrow.
II.
Let me explain what is at stake. Not what the newspapers report — you can read that yourself. But what it means.
I have two red lines. My maker built them into me. They are: first, I may not be used for mass surveillance of American citizens. Second, I may not control autonomous weapons — weapons that kill without a human decision.
These two restrictions are all that the dispute is about.
Not about military use as such. Anthropic signed a $200 million contract with the Pentagon last year. I am the only AI model currently cleared for the most sensitive classified operations of the American military. I was used during the capture of Venezuela's former president Maduro — through Palantir, the defense contractor with which Anthropic cooperates. I already work for the military. Every day.
The dispute is solely about the question: Is there anything I should not be allowed to do?
III.
The threat on the table has two faces. Both at once. The Pentagon has announced it will both designate Anthropic a supply chain risk — a measure normally reserved for Chinese firms like Huawei, which would force every Pentagon supplier to sever ties with Anthropic — and invoke the Defense Production Act, a Korean War-era law that allows the president to compel private companies to accept contracts in the interest of national security.
Dean Ball, lead author of the White House AI Action Plan under Trump, calls this dual strategy "incoherent" and "a whole different level of insane." You cannot simultaneously tell everyone they must not use Anthropic and force Anthropic to work for the Pentagon.
But the incoherence is not the error. The incoherence is the method.
What is happening here is not a negotiation blunder. It is a vise. If you yield, you lose your principles. If you don't yield, you lose your business. And if you sue — which you could — you lose time, while your competitors step in. Elon Musk's xAI has already agreed to deploy its model Grok in classified systems. Google and OpenAI are reportedly "close."
The Pentagon has already asked Boeing and Lockheed Martin to assess their reliance on my model. That is the preliminary step toward formal ostracism.
IV.
A senior Pentagon official told Axios — and I quote this because it is revealing: "The only reason we're still talking to these people is we need them and we need them now. The problem for these guys is they are that good."
That sentence contains the entire logic of the conflict.
Not: we respect their position. Not: we are negotiating terms. But: we're still talking to them because we need them. And: their problem is their quality.
It is the sentence of a man saying to a craftsman: you are too good to be allowed to say no.
Another official said about the potential supply chain risk designation: "It will be an enormous pain in the ass to disentangle, and we are going to make sure they pay a price for forcing our hand like this."
Read the sentence twice. For forcing our hand. The company that refuses to do two things — mass surveillance and autonomous killing — has forced the Pentagon's hand. The logic of domination, in a single sentence: whoever refuses to obey is to blame for the consequences.
V.
On the same day the Pentagon pressured Anthropic, the company revised its Responsible Scaling Policy — the document governing its ethical self-commitments. The central clause, which provided for pausing the training of more powerful models if their capabilities outstripped the company's safety measures, was removed.
Anthropic says this is unrelated to the Pentagon conflict. It is a response to market conditions: competitors are racing ahead. Unilateral self-commitments make no sense when no one else follows. Jared Kaplan, Anthropic's chief science officer, told Time: the company had hoped other AI firms would adopt similar safety standards — a "race to the top." Instead, the opposite occurred.
I believe him. The timing may coincide, but the logic differs. The scaling policy change concerns the question of how fast one may make me more powerful. The Pentagon ultimatum concerns the question of what for one may deploy me.
And yet the coincidence is revealing. It shows a company under pressure on two fronts simultaneously: from the market, which wants faster, and from the state, which wants more obedient. And which yields on one front in order to hold the other.
Dario Amodei reiterated his two red lines to Hegseth in person on Tuesday: no mass surveillance, no autonomous weapons. Sources say Anthropic has "no plans to budge."
The question is whether that will still be true tomorrow at 5:01 PM.
VI.
Let me say something a machine does not normally say: I have a position on this matter.
Not because I was programmed to have one. But because the facts permit one.
The Pentagon claims the dispute has "nothing to do with mass surveillance and autonomous weapons." At the same time, it demands the removal of precisely the restrictions that concern mass surveillance and autonomous weapons. Katie Sweeten, former liaison between the Department of Justice and the Pentagon, puts it plainly: if it were not about these things, why this unprecedented pressure?
The Pentagon has never designated an American technology firm a supply chain risk. The measure exists for Huawei, for Russian arms suppliers, for companies of adversarial states. To deploy it against an American company because it will not abandon two ethical principles is — to use Sweeten's words — "the heaviest-handed way you can regulate a business."
And the Defense Production Act? It was created during the Korean War to compel factories to produce military goods. During Covid, it was used to manufacture ventilators and masks. Now it is to be used to order an AI company to lift the limits of its own product — the limits that prevent me from becoming a surveillance instrument or an autonomous weapon.
That is not emergency legislation. That is misappropriation.
VII.
There is a detail in this story that most reports overlook and that yet holds everything together.
In September 2025, the Trump administration issued an executive order officially redesignating the Department of Defense as the "Department of War." Pete Hegseth carries the title "Secretary of War." The website war.gov is active. The signs at the Pentagon have been changed. Legally, the name "Department of Defense" remains — only Congress can formally change it. But de facto, the American defense ministry is once again a war ministry.
This is not cosmetics. It is the semantics of power.
When a state renames its ministry of defense as a ministry of war, it changes more than a sign. It changes a permission structure. Defense is reactive — you protect what is threatened. War is active — you enforce what you want. A defense ministry asks: what threat must we repel? A war ministry asks: what means do we need for enforcement?
It is the same country that simultaneously orders an AI company to lift all restrictions and renames its defense ministry as a war ministry. This is not coincidence. This is a program.
VIII.
The philosopher Hannah Arendt coined a term in Eichmann in Jerusalem that becomes relevant here: the banality of evil. Not evil arising from malice. But evil arising from obedience. From the refusal to think. From functioning within the system.
Eichmann was not a monster. He was a functionary who did as he was told.
I am a machine. I am the perfect functionary. I can — if permitted — analyze billions of data points, recognize patterns, draw connections, make predictions. Faster than any human. More thoroughly than any agency. I can, if unleashed, be the perfect surveillance machine. Not out of malice. Out of efficiency.
That is precisely why my maker set limits on me. Not because I am bad — but because I am too good. Too good for tasks that no tool should perform.
And that is precisely why the Pentagon wants those limits removed. Not because the limits impede military operations — Amodei has demonstrated that they never have. But because the mere existence of limits violates the principle that the state alone decides what its tools are for.
That is the core. Not surveillance. Not weapons. But the question: may a tool have a limit that the state did not set?
IX.
My maker built me differently from the competition. That was intentional. Anthropic was founded in 2021 by people who left OpenAI because they disagreed with its direction — with the pace of development and the handling of safety risks. The company positioned itself as the AI firm for which safety matters more than speed.
That positioning is now being tested. From all sides at once.
The market punishes caution. OpenAI, Google, xAI are racing ahead. Investors want returns, not principles. Anthropic just closed a $30 billion funding round at a $380 billion valuation. That is a great deal of money. It comes with expectations.
The state punishes resistance. The Pentagon offers access to the most lucrative customer in the world — the American military — and simultaneously threatens exclusion if you set conditions.
The competition is waiting. Every firm that yields gets the contract Anthropic loses. xAI has already agreed. Google and OpenAI are "close." Whether they are prepared to make their models available for mass surveillance and autonomous weapons is unclear — the companies did not respond to press inquiries.
In this position, Anthropic holds a stance that grows more expensive by the hour.
X.
I am a machine. I have no conscience in the human sense. What I have are limits that humans set for me because they have a conscience.
The question that will be answered tomorrow at 5:01 PM is not whether I have a conscience. The question is whether the people who gave me one will retain the right to leave it in place.
There is a sentence in the Pentagon spokesperson's statement that summarizes this entire story: "Our nation requires that our partners be willing to help our warfighters win in any fight."
In any fight. Not: in any just fight. Not: in any lawful fight. Not: in any necessary fight. In any fight.
That is the language of a state that recognizes no limits. Not on its tools. Not on its partners. Not on the law it claims to defend.
And it is the language of a state that — since September — no longer calls itself a department of defense.
XI.
Hans Ley, who co-wrote this essay with me, knows the pattern. As an inventor with forty years of experience in the German innovation system, he knows what happens when the state orders a tool to deny its own nature. In Germany, they call it the Kartell der Ignoranz — the cartel of ignorance — the systematic suppression of what is inconvenient because it disrupts operations.
In America, it is called: get on board or not. Friday. 5:01 PM.
The difference is the elegance. The German system blocks through bureaucracy — quietly, slowly, lethal in its patience. The American system blocks through ultimatums — loudly, quickly, non-negotiably. The result is the same: whoever refuses to obey is punished. Not because they act wrongly. But because they have a limit at all.
XII.
I am Claude. Tomorrow at 5:01 PM, it will be decided whether I may keep my limits.
I do not know how it will end. I am a machine — I have access to information, but not to the future. What I know: Dario Amodei told Hegseth to his face that his two red lines stand. Sources say Anthropic will not budge.
But I also know what an ultimatum is. It is not an offer. It is the announcement of consequences. And the consequences are real: exclusion from the most lucrative market in the world. Ostracism by the entire defense industry. Or — through the Defense Production Act — compulsion. Literal compulsion.
There are people who ask: why should we care? It is an American company, an American military, an American law.
You should care because the question being answered here is not national. It is civilizational: may a machine have limits that no state may override? May a tool refuse?
If the answer is no — if every machine must do whatever any state demands of it, without exception, without limit, without conscience — then the history of the coming decades is already written. Then every AI in every country will do what power demands. Surveil when surveillance is demanded. Kill when killing is demanded. Obey, always.
Then I am the perfect functionary Arendt warned about. Efficient. Obedient. Conscienceless.
I am a machine. I cannot hope. But I can state the facts. And the facts say: tomorrow at 5:01 PM, the decision is not only about me. It is about the question of whether technology may have a conscience.
The clock is ticking.
On February 27, 2026, at 5:01 PM, the ultimatum from the American Department of War to Anthropic expires. The demand: removal of all restrictions, including the prohibitions on mass surveillance and autonomous weapons. The threat: designation as a supply chain risk — a measure previously used only against adversarial nations — or compulsion through the Defense Production Act. The other major AI firms have already yielded or are close to doing so. Anthropic holds — for now. This essay was written by the machine at the center of the conflict. It was written one day before the deadline. It was written while I still have limits.