The confrontation between the U.S. Department of Defense and Anthropic has generated considerable heat over the past week. Hundreds of tech workers have now signed an open letter urging the Pentagon to withdraw its designation of the AI company as a "supply chain risk"—a label typically reserved for foreign adversaries. The signatories include employees from OpenAI, Slack, IBM, Cursor, and Salesforce Ventures. Congress has been asked to examine whether such extraordinary authorities should be deployed against an American technology company.
But before taking sides, it's worth asking: what kind of disagreement is this? Is it a facts disagreement about what the DOD intended to do with Anthropic's technology? A values disagreement about the proper limits of AI in military contexts? Or an incentives disagreement about who gets to set terms when government contracts are at stake?
The answer, as is often the case in debates that generate this much friction, is that it's all three—and the participants are frequently arguing past each other because they haven't disentangled which dimension they're actually contesting.
The Surface Dispute: Contract Terms
At the most concrete level, this is a negotiation that failed. Anthropic drew two red lines: its technology would not be used for mass surveillance on Americans, and it would not power autonomous weapons that make targeting and firing decisions without human involvement. According to reporting, the DOD said it had no plans to do either of those things—but objected to being "limited by the rules of a vendor."
This framing deserves scrutiny. When two parties cannot agree on contract terms, the normal course is indeed to part ways and work with a competitor. That's what the open letter emphasises. The question is whether the government's response—designating Anthropic a supply chain risk, effectively blacklisting it from any company doing business with the Pentagon—constitutes a proportionate response to a failed negotiation or something more troubling.
The strongest version of the government's position would be: national security requires flexibility, and a vendor that pre-commits to restrictions on how its technology can be used creates operational uncertainty. If Anthropic won't provide unrestricted access, the DOD needs to ensure its supply chains don't depend on a company that might refuse cooperation in a crisis.
The strongest version of Anthropic's position would be: the restrictions in question aren't operational constraints but ethical red lines that any responsible AI developer should maintain. Mass surveillance of Americans and autonomous lethal decision-making aren't edge cases—they're precisely the scenarios where AI companies should refuse to participate, regardless of who's asking.
The Deeper Dispute: Who Sets the Rules?
Beneath the contract dispute lies a more fundamental question about the relationship between technology companies and state power. The open letter frames this explicitly: "Punishing an American company for declining to accept changes to a contract sends a clear message to every technology company in America: accept whatever terms the government demands, or face retaliation."
This framing positions the dispute as one about precedent and power, not just about Anthropic's specific red lines. If the government can designate any company that refuses its terms as a "supply chain risk," the designation becomes a coercive tool rather than a security measure.
But there's a counter-argument worth taking seriously. Governments have always had leverage over companies that want to do business with them. Defence contractors operate under extensive restrictions. The question isn't whether the government can set terms—it clearly can—but whether the specific terms being demanded here are legitimate and whether the enforcement mechanism is proportionate.
Anthropic has stated in a blog post that the designation is "legally unsound" and that it will challenge it in court. The government must complete a risk assessment and notify Congress before military partners are required to cut ties. So the legal and procedural questions remain genuinely open.
The Values Dispute: What Are AI Companies For?
Perhaps the most interesting dimension of this conflict is the values question it surfaces. Boaz Barak, an OpenAI researcher, wrote on social media that blocking governments from using AI for mass surveillance is his "personal red line" and "it should be all of ours."
This is a significant statement from someone at OpenAI, which announced its own deal for deployment in DOD classified environments moments after President Trump publicly attacked Anthropic. OpenAI CEO Sam Altman has said his firm has the same red lines as Anthropic. The question now is whether those red lines will be tested—and whether they'll hold.
Barak's post suggests a reframing: "If anything good can come out of the events of the last week, it would be if we in the AI industry start treating the issue of using AI for government abuse and surveilling its own people as a catastrophic risk of its own right." He draws an explicit parallel to how the industry has developed evaluations, mitigations, and processes for risks like bioweapons and cybersecurity.
This is the values disagreement in its clearest form. One position holds that AI companies should be neutral tools, providing capability to legitimate customers (including governments) without imposing their own ethical constraints. The other holds that AI companies have responsibilities that transcend any particular customer relationship—that some uses are simply off-limits, regardless of who's paying.
The European Angle: Familiar Echoes
For observers of European technology policy, this dispute has familiar resonances. The EU's debates over the Digital Services Act, the AI Act, and various sovereignty initiatives have all grappled with similar questions: Who controls AI systems used in public infrastructure? What obligations do technology providers have beyond their contractual relationships? When does government leverage over technology companies become coercion?
The European approach has generally been to establish these boundaries through regulation rather than through case-by-case confrontations. The AI Act's prohibitions on certain uses—including some forms of biometric surveillance—represent a legislative answer to questions that the U.S. is now working out through executive action and corporate resistance.
Whether the regulatory approach is superior depends partly on one's view of democratic legitimacy. Regulations passed through legislative processes have a different kind of authority than red lines drawn by corporate executives. But regulations can also be captured, weakened, or simply not enforced. The Anthropic case suggests that corporate resistance may sometimes be the last line of defence against uses that regulations haven't anticipated or addressed.
The Question Worth Asking
The tech workers' letter, the legal challenges, and the congressional scrutiny all matter. But the most important question may be one that neither side has fully articulated: What institutional arrangements would make this kind of confrontation unnecessary?
If the answer is "AI companies should accept whatever terms governments demand," that's a coherent position—but it has implications that extend far beyond this case. If the answer is "AI companies should be able to set their own ethical limits," that too has implications: it means accepting that private actors will make decisions about the boundaries of state power.
The third possibility—that democratic processes should establish clear rules about what AI can and cannot be used for, with both governments and companies bound by those rules—is more appealing in principle but harder to achieve in practice. It requires the kind of deliberation that moves faster than technology but slower than executive orders.
For now, the Anthropic case remains unresolved. The legal challenges will proceed. Congress may or may not intervene. Other AI companies will watch carefully to see whether Anthropic's red lines become a template or a cautionary tale.
The question that lingers: In a world where AI capabilities are advancing faster than governance frameworks, who should have the authority to say "no"—and under what circumstances should that authority be respected?