When the Court Said "Orwellian": The Pentagon-Anthropic Ruling and What It Reveals About AI Governance Under Political Pressure

Epistemic status: based on court hearing materials from March 24 and the ruling issued March 26, 2026, court filings from both parties, amicus briefs, sworn declarations, and analytical reports from Lawfare, BISI, Brookings, CSIS, and Georgetown CSET. Causal interpretations are my own, with confidence markers where appropriate. I have no insider access to either side of this conflict.

On March 26, 2026, federal judge Rita Lin issued a 43-page ruling blocking the Pentagon’s attempt to designate Anthropic a threat to national security. "Nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S. for expressing disagreement with the government" . The designation was frozen. Trump’s ban order was blocked. The court found "classic illegal First Amendment retaliation," called the government’s actions "arbitrary and capricious," and concluded they "appear designed to punish".

This post is not a retelling of the case timeline (a detailed chronology is available here). It is an attempt to explain why the conflict took the form it did, what the court’s ruling means for AI governance, and what structural problems this case has exposed.

HOW WE GOT HERE: TWO WORLDS AT ONE TABLE

To understand why a contract dispute became a constitutional crisis, you need to see that the two sides at the negotiating table operated under fundamentally different models of decision-making.

On one side — Anthropic and, more broadly, corporate America. Decades of established modus operandi in dealing with the government: negotiations, committees, compromises, lawsuits. Conflicts are resolved through procedure.

On the other side — an administration that, after its first term, made loyalty the central criterion for staff selection. Brookings documents this as a deliberate strategy: 92% of senior appointees departed during the first term, key figures distanced themselves after January 6, and the administration drew a conclusion — in the second term,appoint people whose loyalty is beyond question Defense Secretary Pete Hegseth is a product of this logic: a former Fox & Friends Weekend co-host, a National Guard captain with combat experience but no senior national security experience , confirmed by the Senate 51–50. His decision-making pattern — public ultimatums, social media announcements treated as policy, simultaneous escalation on multiple fronts — is not unique to Hegseth. It is a style honed in the media environment of cable news, where he spent years as a Fox & Friends Weekend co-host, and it bears a noticeable resemblance to the president’s own approach to governance. It works well for dominating a news cycle. It works poorly when the object of pressure has legal options and structural leverage.

His style is codified in the AI Acceleration Strategy of January 9, 2026 — a document written on the heels of a military operation in Venezuela. The key quote: "the risks of not moving fast enough outweigh the risks of imperfect alignment." The same document mandates "any lawful use" for all AI contracts and models "free from usage policy constraints" [Department of War memorandum, January 9, 2026]. The doctrine preceded the Anthropic ultimatum by six weeks — the conflict was execution of a strategy, not improvisation.

The coercive logic — ultimatum, deadline, punishment — had worked with federal agencies. It appeared to work in Venezuela, where, according to Foreign Policy, the use of Claude through Palantir and the subsequent incident with an Anthropic employee triggered the escalation chain. But this logic collided with the fact that an AI company integrated at the IL-6 level is not a subordinate agency. And not Huawei.

Competitors had already accepted the Pentagon’s terms — OpenAI signed "any lawful use" , xAI agreed unconditionally. The calculation looked sound: financial pressure, investor dependency on government contracts, available coercion tools. But all of these tools — supply-chain designation, the Defense Production Act — were designed for markets with interchangeable suppliers of physical goods. You can order masks from another factory. You cannot replace an IT company with classified operations access and deep infrastructure integration in three days.

On February 24, Hegseth issued an ultimatum. Anthropic refused. And then things unraveled.

WHAT THE COURT REVEALED

The March 24 hearing lasted ninety minutes and effectively became a public autopsy of how the designation decision was made.

Judge Lin framed the problem from the opening minutes: "If the worry is about the integrity of the operational chain of command, DOW could just stop using Claude. It looks like defendants went further than that because they were trying to punish Anthropic" .

Two moments effectively demolished the government's position in the courtroom.

First: the Pentagon disavowed its own Secretary. DOJ attorney Eric Hamilton told the court that Hegseth’s post on X — "effective immediately, no contractor may conduct any commercial activity with Anthropic" — had no legal force. Just a "social media announcement." Lin responded: "Pretty surprising... obviously the statement is front and center in this lawsuit" . The Defense Secretary wrote "this decision is final," and his own lawyers told the court to ignore it. For anyone tracking this administration’s decision-making style, the pattern is familiar: impulsive public statement, followed by legal retreat.

Second: the kill switch argument collapsed. Hamilton repeated the Pentagon’s central claim — that Anthropic could "sabotage or subvert IT systems." Lin responded: "What I’m hearing from you is that it’s enough if an IT vendor is stubborn and insists on certain terms and asks annoying questions, then it can be designated as a supply chain risk. That seems a pretty low bar" . Thiyagu Ramasamy, Anthropic’s Head of Public Sector (six years at AWS managing government AI deployments), testified under oath that Claude operates inside an air-gapped system with no remote access, no kill switch, and no ability to push updates without the Pentagon’s knowledge and action . This argument was never raised during negotiations — it appeared for the first time in the government’s court filings.

And there was one more detail the judge included in her ruling: Under Secretary Michael’s email of March 4 — the day after the designation was finalized — in which he wrote to Amodei: "I think we are very close here" . Two days later, the same Michael publicly stated there were no negotiations. A week later — "there’s no chance." The judge wrote directly: the Pentagon’s own records show the designation was motivated by Anthropic’s "hostile manner through the press" — meaning that it went public with the dispute .

Lawfare’s analysis identified the fundamental problem: the statute the Pentagon invoked (10 U.S.C. § 3252) was enacted after a foreign intelligence service compromised classified networks through malware. These are "espionage verbs describing covert, hostile acts by adversaries." The technical characteristics the Pentagon described as threatening — software maintained by an outside vendor with the ability to push updates — describe virtually every piece of software the government uses. If that is sufficient for designation, the secretary could brand any software vendor a security threat over an ordinary contract dispute .

WHERE BOTH SIDES STAND NOW

Anthropic is in its strongest position since the conflict began. The designation is frozen. The ban order is blocked. The company’s strategy is conciliatory: "Our focus remains on working productively with the government" . They are not celebrating publicly — and wisely so: the contract still exists, Claude is still running in Iran, and their best move is to return to negotiations from the position of "we were right, the court confirmed it, let’s make a deal."

The reputational effect has been enormous. Claude reached number one in the App Store, the QuitGPT movement surpassed 2.5 million participants . A company that stood on principle and won — that is a story that sells.

The main risk: a parallel case in the D.C. Circuit under a different statute (41 USC 4713), where the designation has not yet been frozen and two of the three judges on the panel were appointed by Trump . Until both designations are lifted, the relief is incomplete.

The Pentagon is in a difficult position. It lost a round with devastating judicial language. Under Secretary Michael called the ruling a "disgrace DURING A TIME OF CONFLICT" but did not name a single specific error . DOJ is filing an appeal to the Ninth Circuit. Formally, the Pentagon continues to consider Anthropic a supply-chain risk, but the court has barred enforcement. Claude continues to operate in Iran.

Judge Lin emphasized: the Pentagon has every right to stop using Claude and switch to another provider. It simply cannot punish a company for its public position. This leaves an exit: quietly negotiate acceptable terms and frame it as "we always wanted to work together." The question is whether the administration’s style will permit a quiet deal.

IRAN AS AN AMPLIFIER

The war is on day 29. The Center for American Progress estimates total costs at approximately $25 billion as of March 26 . CSIS calculated $3.7 billion in the first 100 hours, $11.3 billion by day six, and $16.5 billion by day twelve . The Pentagon has requested an additional $200 billion from Congress . Thirteen American service members have been killed, along with over 1,000 Iranian civilians. Iran rejected the 15-point peace plan, issued a five-point counterproposal including sovereignty over the Strait of Hormuz , and publicly denies it is negotiating at all— though messages are flowing through Pakistani, Egyptian, and Turkish intermediaries. The prospects for a near-term ceasefire appear dim: Iran suspects a trap, having twice before seen Trump approve attacks during talks, and Israel is accelerating strikes on nuclear facilities and arms factories in case a ceasefire is declared. Trump has twice extended his deadline for strikes on Iranian energy infrastructure , but the 82nd Airborne is deploying — preparation for a ground option, or a bargaining chip, or both.

For the Anthropic dynamic, this means the Pentagon is under pressure from multiple directions, and the AI company dispute is an unwelcome additional front. Every week without a clear result in Iran strengthens Anthropic's position — not through anything the company does, but because circumstances work in its favor. If the war ends, Anthropic will remain the only company with experience operating military AI infrastructure in real combat conditions. That advantage cannot be replicated.

AN UNPRECEDENTED COALITION

The scale of support for Anthropic reveals who was frightened by the designation.

Microsoft filed an amicus brief calling the designation "unprecedented overreach" — having invested $5 billion in Anthropic, it was also protecting its own investment. 150 former judges warned of constitutional violations . 22 retired generals and admirals stated that Hegseth’s actions constituted "retribution" and that abrupt tool changes could "harm troops in theater" . Nearly 50 employees of OpenAI and Google DeepMind — competitors who directly benefit from Anthropic’s exclusion — supported the company, writing that the Pentagon had acted "recklessly" . A coalition of EFF, FIRE, and the Cato Institute called the designation "the Pentagon’s temper tantrum" .

On the Pentagon's side — only DOJ, by duty of office. Not a single company, organization, or former official filed a brief supporting the designation.

This is telling. The designation was conceived as a tool of pressure against Anthropic. It became a signal to the entire industry: if Anthropic today, then any company tomorrow that refuses the administration's terms. The industry responded not with solidarity toward Anthropic, but with self-preservation.

THE STRUCTURAL ASYMMETRY: A WIDER VIEW

Sarah Kreps of Cornell’s Tech Policy Institute puts it this way: "What began as a contractual disagreement has evolved into a broader debate about the relationship between governments and the companies building the most advanced AI systems" . I would add: it evolved not organically, but because a specific decision-making style — impulsive, public, built on ultimatums — turned a contract dispute into a constitutional crisis.

The Pentagon could have simply declined to renew the contract and quietly transitioned to another provider. Judge Lin stressed this point. Instead, it applied an instrument designed to combat foreign spies against a company it had nearly reached agreement with the day before. This is not an accident — it is a predictable outcome of a system in which decision-makers are selected for loyalty and apply coercive logic to objects that do not yield to coercion.

For comparison, it is worth looking across the Pacific. Under China’s Military-Civil Fusion strategy, eleven of the fifteen largest AI suppliers to the PLA are state-owned enterprises or affiliated defense research institutes . The 15th Five-Year Plan institutionalizes AI Plus, embedding military requirements into civilian R&D from the outset . A conflict like Anthropic-Pentagon is structurally impossible there — not because Chinese companies are more compliant, but because the relationship is designed differently.

The paradox noted by Fortune sharpens this asymmetry: Anthropic and OpenAI have publicly accused Chinese labs of distilling their models; unrestricted versions are already available to the PLAg . American companies restrict their own military while adversaries train on pirated versions of the same technology with no restrictions whatsoever. By pursuing Anthropic, Washington risks weakening its strongest AI company at the moment it needs it most.

WHO BENEFITS

Anthropic's win benefits: AI companies (precedent: the government cannot punish firms for their public position on AI safety), civil society (the First Amendment applies to executive overreach in AI), Democrats (a midterms narrative for 2026), and Anthropic's investors (stability for $5+ billion in investments).

A Pentagon win would have benefited: the administration (a precedent for compulsion), advocates of strong executive authority, Anthropic's competitors (xAI, partially OpenAI) — and, paradoxically, China, because weakening America's strongest AI company during a technology race is a gift to Beijing.

WHAT THIS MEANS

Anthropic won a round, but not the war. The D.C. Circuit, the Ninth Circuit, a possible appeal — the case will drag on for months. Meanwhile, Claude continues to run in Iran, and every day of operation is another argument for indispensability.

The Pentagon lost a round but preserved the right to switch providers — the judge emphasized this. The question is whether the transition will be quiet and rational or public and confrontational. The administration's style so far points to the latter.

The war in Iran makes everything harder. A ceasefire would ease pressure on both sides and create space for a quiet deal. A prolonged war keeps Anthropic indispensable, and every week strengthens its position.

But one conclusion is already final: the court has said that national security instruments cannot be used to settle contract disputes with American companies. This precedent will shape every future negotiation between an AI company and a government — in any country.

And one more question this case has raised but not answered: Anthropic's red lines protect Americans — not people in general. The phrase "mass domestic surveillance of Americans" leaves 7.7 billion people outside the perimeter. In the context of an operation where Claude is identifying targets in Iran, this distinction ceases to be abstract. But that is a topic for a separate post.

To be clear: Anthropic should not be idealized in this story. Its red lines are narrow and self-serving in a precise way — they protect the company from specific reputational risks (autonomous weapons, domestic surveillance) while leaving the rest of Claude’s military applications untouched. The phrase "any lawful use" is quietly becoming the default standard for defense AI procurement, and Anthropic’s own position is closer to that standard than its public framing suggests.
But there is a deeper problem that this case has surfaced without resolving. The entire dispute — whether the Pentagon accepts Anthropic’s red lines or overrides them — is ultimately irrelevant to the 7.7 billion people who are not Americans. For them, the question of who wins this contract negotiation changes nothing. Under either outcome, they remain within the category of permissible targets. Anthropic’s red lines were never drawn around them.
The deeper question — whether an AI company should have the right to set usage boundaries at all, or whether that authority belongs solely to democratic institutions — remains open. But so does a prior question: whose protection are those boundaries actually designed for? That is what I am examining next.