The Line That Moved

Two things happened in the same news cycle. The Trump administration ordered every federal agency to stop using Anthropic's tools, because Anthropic would not lift the restrictions in its usage policy that limit how its models can be applied to weapons systems and certain military operations. And within hours, hours, OpenAI announced a classified partnership with the Pentagon, with Sam Altman putting his name directly on the announcement.

The ordinary way to cover this is as a business story. Anthropic lost a customer. OpenAI won one. The federal AI market is large and contested and the frontier labs are now openly competing for it. All true, and all beside the point.

The federal AI market is large and contested and the frontier labs are now openly competing for it.

What actually happened is that a line moved. Not a line in a contract. A line about what the people building these systems are willing to refuse, and under what pressure, and for how long. I want to describe what moved, who moved it, and who will pay for the fact that it moved, because I don't think the consequences of this week will be felt primarily by the two companies involved. They will be felt by the people on the other end of the systems that get built now that the refusal has been punished and the compliance has been rewarded.

What the refusal actually was

Anthropic's Usage Policy has, for most of the company's history, carved out specific military and weapons-related applications as prohibited or restricted. The exact wording has shifted, the company has layered on carve-outs for certain government uses, cybersecurity applications, intelligence analysis, but the core has held: the models may not be used to develop weapons, to target individuals for kinetic action, or to operate offensive systems without human control. When the administration pushed for unrestricted access across agency use cases, including applications that would have required Anthropic to waive those restrictions, Anthropic declined. The administration responded by ordering agencies off the product.

I want to be careful here. Anthropic is not a moral hero of this story. It is a company that has argued, loudly and often, that it is a safety-first frontier lab, and its usage policy is part of that argument. The policy is a product of both conviction and positioning; you cannot cleanly separate them. Anthropic has accepted defence-adjacent work before. It has a CIA-adjacent relationship through its Palantir partnership. It has built for Anduril. Its line is not "no military", it is a specific, negotiated line about what kinds of military use its models can be put to.

But a negotiated line is still a line. And this week, the administration tested it, and Anthropic held, and the administration punished the holding. That sequence, test, hold, punish, is the thing to pay attention to.

What OpenAI's announcement did

Within hours of Anthropic's ouster, OpenAI announced its classified Pentagon partnership. The timing is not coincidence. I don't think anyone, including OpenAI, is pretending it is. Altman made the announcement himself, which tells you how the company wants it read: not as a quiet commercial deal signed in the ordinary course, but as a statement of posture.

The posture is: we will do the work that the other lab would not.

OpenAI has been moving in this direction for more than a year. In January 2024 the company quietly removed the blanket prohibition on "military and warfare" applications from its usage policy, replacing it with more specific language about weapons development and harm.¹ It has since taken Pentagon contracts, built with Anduril, and accepted a $200m Defense Department agreement in 2025.² The classified partnership announced this week is not a pivot. It is the culmination of a trajectory that has been legible for some time.

What is new is the explicit framing of this deal as filling a gap created by another lab's refusal. That framing is what makes the week consequential. It converts a refusal, Anthropic's, into a competitive liability. It tells every other AI company, current and future, that maintaining military-use restrictions is a thing you will be punished for, and abandoning them is a thing you will be rewarded for. The structure of the incentive is now visible to everyone.

The market for compliance

I've been thinking about a framing Timnit Gebru and others have used for a while, which is that AI ethics inside large labs is, functionally, a form of marketing that becomes operationally binding only when it doesn't conflict with revenue. When it does, the ethics gets rewritten. What's different now is that the conflict has become forced rather than latent. The federal government has moved from being one of many customers whose preferences had to be balanced against others, to being a customer powerful enough to require that the preferences of other stakeholders, safety teams, ethics boards, the broader public, be subordinated to its demands.

This is what "the line moved" means in practice. It does not mean that Anthropic will now change its policy (though it might, under continued pressure). It means that the market has acquired a new mechanism for disciplining labs that maintain restrictions. The mechanism is: the federal government will route around you, and a competitor will take the contract, and the loss of revenue will be attributed internally to your ethics team, and the internal politics of your company will shift accordingly.

I find it hard to believe that any of the frontier labs will hold firm lines against militarisation over a five-year horizon under this pressure. The incentive structure is now too clear. The labs that hold out will lose federal revenue and, over time, lose the talent and capital that follows federal revenue. The labs that don't will win. This is not a prediction about intentions, I expect many people inside Anthropic genuinely believe in the restrictions they have fought for. It is a prediction about incentive gradients, and the historical track record of incentive gradients versus intentions inside high-growth companies facing existential revenue pressure.

Who pays

Here is where I want to stop talking about companies and talk about people.

A policy line that restricts how an AI model can be used in military applications is not an abstraction. It is, concretely, a set of constraints on what the model will help a user do when the user is, say, an intelligence analyst building a targeting package, or an operator running a drone mission, or a contractor developing a new weapons system. Remove the restrictions, and the model helps. Keep the restrictions, and the model refuses, or degrades, or routes the request to a human reviewer.

The people on the other end of those requests are not American procurement officers. They are the people who get targeted. They are the people in the building that gets hit. They are the people whose metadata patterns fit the classifier. When an AI model goes from "will not help with this" to "will help with this," what changes materially is the cost, speed, and scale at which lethal or coercive action can be produced. Every serious study of autonomous and AI-assisted military systems says the same thing: the binding constraint on operational tempo has been human cognitive bandwidth, and AI's role is to relax that constraint. Relaxing it means more actions per unit time. More actions per unit time, against a targeting apparatus that already has documented false-positive rates in the double digits for some classes of target, means more wrong people affected.³

I am aware that this is the part of the argument where techno-optimists reach for the rejoinder that AI also makes targeting more accurate, reduces collateral damage, enables proportionality assessments that humans can't do under time pressure. I take that rejoinder seriously. It is not empty. In some applications it's probably even true. But the relevant question is not whether AI can, in principle, reduce harm in military applications. It is whether the specific deployments that the specific contracts being signed this week are specifically for will, in practice, reduce harm, and whether the people most affected by those deployments have any voice in whether they happen.

The answer to the second question is: no, they do not. The people who will be on the receiving end of AI-assisted military action by US agencies are, overwhelmingly, not US citizens, not represented in US democratic processes, and not consulted about US procurement. The consent problem here is not subtle. It is the oldest consent problem in the history of military technology, which is that the people who decide to deploy a weapon and the people the weapon is deployed against are never the same people, and almost never even in the same political community.

What's specific to this moment is that the domestic institutions that have historically provided some friction against expansion of military AI use, ethics boards, usage policies, researcher protest, congressional oversight, are being systematically routed around or punished. The Anthropic ouster is an instance of the punishment. The OpenAI partnership is an instance of the routing-around. The result is a procurement environment in which the friction has been removed, and the only remaining constraint is what the labs themselves are willing to refuse. And we've just watched, in real time, what happens to a lab that refuses.

The domestic version

I want to widen the frame one more step, because the international story is not the only one.

The same administration that ordered agencies off Anthropic's tools has, over the past fifteen months, expanded the federal government's use of AI systems for domestic applications that directly affect US residents, immigration enforcement, benefits adjudication, surveillance of protest activity, identification of federal employees for removal based on ideological screening. These deployments do not require waivers of the military-use restrictions in anyone's usage policy. They require, instead, waivers of restrictions on surveillance, on discriminatory use, on use against specific protected categories of person. Those restrictions also exist, in various forms, in various companies' policies. They are also under pressure.

The Anthropic–OpenAI swap is the most visible instance this week of the government demonstrating its willingness to punish labs that hold lines. It is not the only instance, and the military applications are not the only applications where the lines matter. If the pattern continues, and the incentive structure now says it will, we should expect to see similar test-hold-punish sequences across the rest of the usage-policy surface. Surveillance restrictions. Restrictions on use in hiring and benefits decisions. Restrictions on political and ideological profiling. Each of those will be tested. Some will be held. The ones that are held will be punished. The ones that are dropped will be rewarded.

The people who will pay for those drops are, in each case, specific and identifiable. Asylum seekers whose claims are adjudicated by a model with no restrictions on flagging them for adverse outcomes. Benefits recipients whose eligibility is determined by a system with no restrictions on demographic inference. Federal workers whose political activity is surfaced by a model with no restrictions on ideological classification. Each of those populations is on the receiving end of a usage-policy line that is currently under pressure, and will, if the pattern holds, eventually move.

What the week was

So the week was not a business story. It was a governance story, and it was a consequences story, and the consequences will not land on the two companies involved. They will land on people, mostly people without standing in the political system that is producing the policies, who will experience the softening of usage-policy lines as, concretely, more AI-assisted decisions made about them with fewer constraints on how those decisions can be produced.

I want to end where I started, which is the question of what refusal is for. Anthropic's refusal this week did not prevent the Pentagon from getting what it wanted. The Pentagon got what it wanted within hours, from a competitor. In that narrow sense the refusal was ineffectual, and some commentary this week has been in that register, why bother holding a line that moves around you.

I think that's the wrong read. A refusal that fails commercially is not the same as a refusal that fails morally, and the continued existence of lines that some actors will not cross is the precondition for any future regime in which those lines get institutionalised rather than left to the discretion of individual firms. The hard lesson of this week is not that Anthropic was wrong to refuse. It is that a regulatory architecture in which the only constraint on military AI use is the usage policy of whichever lab happens to hold the contract is not an architecture. It is a temporary arrangement, dependent on the unstable commercial preferences of a handful of companies, and this week we saw how unstable.

The line moved because there was nothing holding it in place except the willingness of one company to lose a customer. That turns out to be not very much.

Footnotes

¶