Anthropic's Red Lines Are No Substitute for Public Law

This post is part of a series of student essays produced in collaboration with the Berkman Klein Center for Internet & Society at Harvard University. Read more in the series here.

As a student in Beijing, I enjoyed the mundane conveniences that facial recognition technology provided. When I went to watch the China Open, the country’s premier tennis tournament, a cursory scan of my face produced directions to my seat. At Daxing International Airport, a small kiosk scanned my face and displayed my gate, seat, and boarding time. On street corners, my face unlocked refrigerated vending machines from which I drank cold jasmine tea while several yuan were deducted from my AliPay account. In China, more than in the United States, artificial intelligence is an ambient feature of the commercial ecosystem, a lubricant for the minor frictions of everyday transactions.

Whether or not you find these applications of AI benign or troubling, notice that none were implemented by institutions that the public could vote out. The United States is now drifting toward its own version of undemocratic AI governance: today, the most consequential limits on the American government’s AI use appear not in statutes or regulations, but in the contractual terms of a single private company. Such contractual provisions, however well-intentioned, are insufficient to protect the American people against AI harms. They are undemocratic in origin, brittle in practice, and, most dangerously, a de facto substitute for the more robust protections that only public law can provide.

Harmful technologies seldom announce themselves; they arrive first as benign conveniences. Browser cookies, introduced in 1994 to save items in users’ carts between shopping sessions, later enabled consumer surveillance and targeted advertising. Facebook’s Like button began in 2009 as a way to show appreciation for friends’ posts; later, its descendants became the engine of a teenage mental health crisis. Aza Raskin, who created the infinite scroll in 2006 to create a more seamless user experience, deeply regrets an invention that now wastes 200,000 human lifetimes daily. Convenient technology can embed itself in daily life, only revealing its darker applications after millions have come to depend on it.

Congress’s abject failure to remedy, much less anticipate, these technological harms has led to an unreliable regime of industry self-regulation. Anthropic’s Responsible Scaling Policy, which governs when its models are deemed too dangerous to deploy, has been criticized on exactly this ground. OpenAI has received similar criticism for changes to its mission statement and for dissolving its “mission alignment” team. In 2023, seven leading AI companies voluntarily pledged at the Biden White House to conduct safety testing before releasing powerful models. By 2025, most of the seven had quietly walked them back.

But we are now seeing a different type of industry regulation—not of itself, but of its government clients. Call it “reverse regulation”: private firms dictating the terms under which public institutions may wield a critical technology. Generally speaking, attaching such use restrictions is legal and commonplace. But we should hesitate when private firms use their leverage to dictate restrictions concerning the most fundamental questions of privacy and national security.

The standoff between Anthropic and the Department of Defense (DoD) is the starkest example. Anthropic, whose Claude model has become deeply integrated into military and intelligence systems, refused to remove two contractual red lines: that Claude not be used for mass domestic surveillance and that it not power autonomous lethal weapons. Secretary of Defense Pete Hegseth responded by designating Anthropic a supply chain risk (a label previously reserved for foreign firms like Huawei) and threatening to invoke the Defense Production Act.

The public reaction split predictably. On one hand, many concerned about AI safety and civil liberties felt that Anthropic was “absolutely heroic.” On the other hand, many national security hawks agreed with Hegseth’s assessment that it was “a master class in arrogance and betrayal” and “a cowardly act of corporate virtue-signaling.” Both camps are missing a deeper democratic problem.

Set aside, for a moment, whether you agree with Anthropic’s substantive positions. (I do.) The prospect of the Trump administration employing AI for domestic surveillance is harrowing, and this essay is not an endorsement of Hegseth’s response; designating Anthropic a “supply chain risk” is a brazen abuse of power that many amici, and even OpenAI, denounced. The question is not whether Anthropic’s red lines are sensible, but rather whose role it is to set them. In a democratic republic, it must be the people and their elected representatives—not a private company’s boardroom, however wise or well-intentioned.

When Anthropic attaches policy conditions to defense contracts, it makes a public policy decision that Congress otherwise might: determining, on behalf of almost 350 million Americans, what the United States military may and may not do with a critical technology. It is, of course, within Anthropic’s legal rights to set terms for its services. But unlike Congress, Anthropic’s leadership is unelected; it holds no constitutional mandate; there is no guarantee that its normative views will align with the public’s; and the public has no mechanism to hold it accountable for its stipulations to DoD.

Even in the absence of congressional action, the executive branch could promulgate regulations preventing the use of AI in autonomous weapons. The DoD’s failure to do so is itself a piece of information on which voters might base electoral decisions. This particular standoff did not suppress public attention; to the contrary, it produced a spectacle. But had DoD accepted Anthropic’s stipulations, voters would have been deprived of information about the Trump administration’s approach to military AI.

The problem with governing by contract is not merely that it is undemocratic, but also that it produces dangerously brittle protections of fundamental civil liberties. If Anthropic’s CEO or board changes, its contractual language might be quietly revised, leaving the American public without recourse. Anyone under the illusion that the company’s voluntary commitments are durable need only look at its stripped-down Responsible Scaling Policy Version 3.0. Anthropic’s Jared Kaplan admitted, “We didn't really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments…if competitors are blazing ahead.” Americans who believe they are protected from AI-driven surveillance are, in reality, protected only by the Anthropic leadership’s policy preferences. Worse still, those preferences are about to answer to public shareholders: Anthropic’s June 2026 IPO filing, at a valuation near one trillion dollars, makes maintaining a costly red line precisely the kind of choice that fiduciary duty disfavors.

More insidiously, private red lines displace the political urgency for democratic action in the long run. When private red lines do the work of public law, the political pressure that might produce more robust legislation may abate. Had DoD accepted Anthropic’s terms, they would have taken effect quietly. Yet, when a contractual provision quietly limits what a federal agency may do with a critical system, no one protests, no legislator convenes a hearing, and no vote is taken. The constituency that might ordinarily demand legislative action has less to gain by doing so.

Setting aside democratic concerns, contractual enforcement may well fail in practice. Anthropic almost certainly cannot monitor how the DoD uses Claude in certain classified environments. And if the DoD breached the contract, federal courts would be deeply reluctant to enjoin a sitting president’s use of AI in military operations. Although the red lines exist on paper, the enforcement mechanism may be only the government’s good faith, which is precisely what Anthropic does not trust.

It is worth distinguishing Anthropic’s actions from conventional corporate political activity. When Exxon, say, lobbies Congress to favor fracking over environmental regulation, it advocates its narrow interests above the public’s. But Anthropic is claiming to do the opposite: It is acting in the public good to its own financial detriment. Amodei has written of his willingness to incur costs that “cut into our margins” because it “is the right thing to do.” That may sound nobler, but precisely because Anthropic claims to act for everyone rather than for itself, it supplants the democratic process more completely than any lobbyist does.

Many dissenters are pragmatic: Amodei’s red lines, they contend, are a sensible response to a dysfunctional Congress, and waiting for that Congress is fanciful. But the illusion of protection is more dangerous than its visible absence. History bears this out. In the wake of the Watergate scandal, the Church Committee exposed rampant domestic surveillance by the CIA and FBI. The public reacted, and Congress introduced FISA in 1978, which was amended in 2008 after yet more domestic surveillance by the NSA was uncovered. Durable protections of civil liberties require public exposure, public outrage, and public law, even if they are born of a slow and imperfect process. Anthropic’s red lines are a substitute for that process, not a step toward it. Anthropic itself seems to recognize this point: in June 2026, when the government forced it to disable its newest models, the company insisted that such limits belong in “a statutory process that is transparent, fair, clear, and grounded in technical facts,” which is to say, in public law.

It is tempting, when facing an administration as reckless as the current one, to cheer any institution that stands in its way. But the right response to democratic dysfunction is not to cede governance to private corporations. The more consequential a question, the more deserving it is of a democratic answer. Correct though they are, these red lines are not Anthropic’s to draw.