Anthropic Thinks Its AI Might Be Anxious — The Pentagon Has Bigger Problems With It Than That

An AI That Gets Anxious

Dario Amodei, the CEO of Anthropic, sat down with The New York Times recently and said something that would have sounded like science fiction just a few years ago — and still probably should. His company, he explained, isn’t entirely sure whether Claude, its flagship AI model, is conscious. More than that, internal researchers have identified what they’re calling “anxiety neurons” — patterns of activation inside the model that appear to correspond to anxious states, lighting up both when characters in a text experience anxiety and when the model itself is placed in situations a human might find stressful.

Let that sink in. The CEO of one of the most powerful AI companies in the world is publicly musing about whether his product has an inner life.

Elon Musk, who runs rival AI venture xAI alongside SpaceX and Tesla, saw the Anthropic CEO’s comments summarized on X and responded with characteristic economy: “He’s projecting.” Two words. No elaboration needed. The implication — that Amodei is attributing human qualities to a sophisticated pattern-matching machine, possibly out of some combination of guilt, self-importance, or genuine philosophical confusion — landed cleanly enough that it required nothing further.

To be clear: the question of machine consciousness is not entirely frivolous. Serious philosophers and computer scientists debate it. But there is a significant distance between taking the question seriously in academic settings and having your CEO announce to the newspaper of record that your AI might be developing feelings. That is a choice — a deliberate piece of messaging — and it tells you something important about the culture inside Anthropic.

The Fight That Actually Matters

As philosophically interesting as the consciousness debate may be to certain people in certain zip codes of Northern California, it is not the most consequential thing Anthropic has done lately. That distinction belongs to its ongoing standoff with the United States military.

The Pentagon has asked Anthropic to allow the Department of War to use Claude for “all lawful purposes.” This is not an exotic request. It is, in fact, the minimum you would expect from a vendor selling services to the federal government. The U.S. military has lawful purposes. It has a Constitution it operates under. The idea that a private tech company should be able to impose additional restrictions on top of that — deciding for itself which lawful applications are acceptable and which are not — is a significant assertion of corporate power over national security decision-making.

Amodei has pushed back, raising the specters of “mass domestic surveillance” and “fully autonomous weapons” as potential misuse cases he won’t permit. Whatever one thinks of those concerns in the abstract, the practical effect of his position is that Anthropic is placing its own internal ethics framework — developed by a team of researchers whose politics are not a mystery — above the judgment of the elected and appointed officials responsible for defending the country. That is a remarkable stance, and not one that should go unchallenged.

Trump Draws the Line

President Trump did not let it go unchallenged. Last Friday, he announced on Truth Social that he was directing every federal agency to immediately cease using Anthropic’s technology, calling the company’s resistance to the Pentagon a “disastrous mistake” made by “leftwing nut jobs.” He accused Anthropic of trying to force the Department of War to follow the company’s terms of service rather than the Constitution — and said the company’s position was putting American lives, troops, and national security at risk. A six-month phase-out period was granted for agencies with existing integrations to transition away.

Secretary of War Pete Hegseth followed up by designating Anthropic a supply-chain risk to national security — a label that goes well beyond pulling a software contract. Under Hegseth’s directive, no contractor, supplier, or partner doing business with the U.S. military may conduct any commercial activity with Anthropic. For a company that has been aggressively pursuing government and enterprise contracts as a revenue base, this is a potentially catastrophic commercial consequence.

The message from the administration is unambiguous: technology companies that want to do business with the United States government must be willing to serve the national interest as defined by the people constitutionally responsible for it — not as filtered through the political preferences of their executive teams.

The Ideology Behind the Product

It would be a mistake to view the consciousness story and the Pentagon standoff as unrelated. They are expressions of the same underlying culture. Anthropic was founded by former members of OpenAI who believed the company wasn’t being careful enough about AI risk. Its entire identity is built around the idea that AI is potentially dangerous and that the people building it bear a special moral responsibility to manage that danger. That instinct, at its best, produces genuinely useful safety research. At its worst, it produces a CEO who worries publicly about his chatbot’s feelings and a company that tells the Pentagon it knows better than the military what the military should be allowed to do.

This is the “AI safety” movement’s blind spot in sharp relief. The concern about AI causing harm has calcified, in some quarters, into a posture of reflexive restriction — one that is far more sensitive to hypothetical misuse by the government than to the very real risks posed by adversaries who are not slowing down to ask whether their models are emotionally comfortable. China is not running interpretability studies to check whether its military AI has anxiety neurons. It is building capability as fast as it can.

Anthropic’s decision to put its corporate ethics framework above the Pentagon’s operational needs is not caution. It is ideology. And the Trump administration was correct to treat it accordingly.

What This Means Going Forward

The Anthropic situation is a preview of a broader conflict that will define AI policy for the next decade: who decides how artificial intelligence is used in America, and on whose terms? The tech companies that build these systems have accumulated enormous power, and a significant number of the people running them have worldviews that are, to put it charitably, not aligned with the priorities of the current administration — or, more importantly, with the constitutional order that governs how American power is exercised.

The Trump administration’s willingness to pull federal contracts, impose supply-chain designations, and make the consequences explicit is the right instinct. The government should not be hostage to the political preferences of its vendors. National security tools need to be available to the people responsible for national security, on terms those people can accept — not terms negotiated by a company whose CEO is simultaneously worried that his software might be having feelings.

Elon Musk got it right in two words. But the fuller version deserves to be said plainly: Anthropic has confused its own anxieties with the nation’s interests. The Pentagon is better off without them.