On September 25, 2026, the U.S. Court of Appeals for the District of Columbia Circuit ruled 2-1 in favor of the Department of Defense’s supply-chain risk determination regarding AI company Anthropic, upholding a ban that continues to prohibit the military and defense contractors from using Claude AI models. The ruling marks the first time that “AI alignment technology” itself has been characterized as a source of government security concern.
The Core Logic of the Ruling: Motive Does Not Matter
The majority opinion was written by Trump-appointed Judge Gregory G. Katsas, joined by fellow Trump appointee Judge Neomi Rao. The central logic of the ruling was strikingly direct: “The Department of Defense had substantial evidence to determine that continued integration of Claude into military information systems by the department or its contractors constitutes a statutory national security risk. As Anthropic itself acknowledges, the company encoded restrictions into Claude to prevent the model from performing tasks that Anthropic does not want it to perform.”
The key lies in the second sentence. The court did not question Anthropic’s intentions—the majority opinion explicitly stated that “we have no reason to doubt Anthropic’s noble intentions in restricting Claude”—but immediately added that the supply-chain risk determination “turns on what Anthropic did, not why it did so.”
“The Department of Defense reasonably worried that Anthropic could manipulate Claude’s design so that it would be unable to perform national security functions that the department regards as contractually authorized and necessary.” — Judge Gregory G. Katsas
The Pentagon does not trust any AI company to have the unilateral authority to decide what the military may use AI for, even if that company’s starting point is safety and ethics.
Two Bans, Two Courts, Two Completely Opposite Conclusions
To understand this ruling, it is necessary to see the case’s dual-track structure. The Pentagon invoked two separate supply-chain statutes, and the corresponding lawsuits were therefore heard separately in courts in two jurisdictions.
One track had already moved in Anthropic’s favor: on August 27, 2026, Judge Lin of the U.S. District Court for the Northern District of California ruled in the corresponding case that the Pentagon’s actions violated the First Amendment—finding that the move was unlawful retaliation against Anthropic’s free expression—and rejected the supply-chain risk determination as “arbitrary and capricious.”
The majority opinion in Washington did not ignore this. The ruling expressly stated that the majority judges had “no disagreement” with the California court’s conclusion, but held that the two cases involved different statutory provisions and could therefore reach different results. The dissent—written by Judge Karen LeCraft Henderson, appointed by President George H.W. Bush—argued that the Department of Defense had exceeded its authority.
Two federal courts issuing diametrically opposed rulings in parallel cases provides the strongest grounds for further appeal. Anthropic spokesperson Danielle Cohen said the company is “considering all options, including further judicial review”—meaning it may seek rehearing en banc before the D.C. Circuit or take the matter to the Supreme Court. The court has delayed the ruling’s immediate effect, giving Anthropic a window to file its request.
Alternatives Are Already in Place, and the Market Landscape Has Shifted
The legal process is still ongoing, but the Pentagon’s replacement plan has largely been completed. In late August 2026, the Department of Defense successively added OpenAI’s ChatGPT Mil and xAI’s Grok for Government to its AI portal, GenAI.mil; Google Gemini had previously gone live on the platform. These three tools are currently available to roughly 3 million Pentagon civilian employees and active-duty service members. Claude is entirely absent from the list.
OpenAI and Google both faced public opposition from employees when signing contracts with the military, for precisely the kinds of military use cases Anthropic had initially refused to accept. Both companies ultimately cited “supporting the U.S. government” as their rationale, overcame internal dissent, and completed the signings. This has placed Anthropic at a double disadvantage in commercial competition: it has lost an important government customer while also creating a clear distinction from its peers on the question of whether to cooperate with the military.
A Distorted Incentive Structure: Stronger AI Safety, Weaker Market Position
The logic of the D.C. Circuit’s ruling objectively creates an abnormal incentive: an AI company that encodes ethical constraints into its model becomes more vulnerable in government procurement than competitors without such constraints. The military’s argument—that “the company could manipulate model design to prevent military uses”—applies precisely to every developer that takes AI safety seriously.
Matt Schruers, president of the technology policy research organization CCIA, stated this directly: “Designating a company as a supply-chain risk is a tool typically reserved for foreign adversaries and must be used carefully, not as a punitive measure.” CCIA, ITI, SIIA, and TechNet had all previously submitted amicus briefs to the court, warning of the potential damage this approach could cause to U.S. technology competitiveness.
The key issue lies at the starting point of the dispute: the Pentagon demanded that contracts include an “all lawful uses” clause, requiring AI developers not to impose restrictions on what the military can do with AI. The two red lines clearly drawn by Anthropic CEO Dario Amodei—autonomous weapons and mass surveillance of U.S. citizens—were viewed by the military as voluntary limitations on the company’s ability to fulfill the contract, rather than responsible safety practices.
The Political Coordinates of the Ruling
The nominating backgrounds of the three judges are not merely labels in this case. The two judges in the majority were both nominated by Trump, while the dissenting judge was nominated by George H.W. Bush. This composition cannot be used to challenge the legal validity of the ruling, but it objectively places the case within the coordinates of the current U.S. political environment: under the federal judicial framework shaped by the Trump administration, the standard for accepting national security arguments is undergoing a systemic drift.
Nor did this tug-of-war occur in a vacuum. Throughout the legal battle, the Pentagon never provided a detailed explanation of how it would gradually replace Claude. The timing of GenAI.mil’s launch—one month before the ruling was handed down—shows that the military had already completed its strategic backup before the legal outcome became clear.
Anthropic’s Position and Its Judgment
For Anthropic, the short-term financial impact of this ruling remains unclear. At the early stage of the supply-chain risk determination, the company said it had lost revenue because customers were waiting on the sidelines, but it has since remained broadly positive about growth and has continued moving toward an IPO. Its performance in the commercial market indicates that the loss of government contracts does not pose an existential threat.
But the significance of the legal precedent is long-term. This ruling is not only about Anthropic; it also concerns the basic framework for how the entire AI industry will work with the federal government in the future: Who gets to define the “lawful uses” of AI? Do private companies have the right to build ethical constraints into their own products that buyers cannot rewrite?
The courts in California and Washington have given completely different answers. That alone shows that these questions are far from settled. Until a final ruling from a higher court emerges, the central contradiction in U.S. AI regulation will continue to evolve around this case. As for the AI companies that have already signed military contracts—before this debate has reached its conclusion, the history in which they chose to stand on the other side will accompany them as well.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接