In September 2026, Dario Amodei proposed an unusual response to the race to develop frontier artificial intelligence. If competition pushes leading laboratories to move faster than safety allows, Washington should let them coordinate. The Anthropic chief executive wants outside evaluators embedded inside AI companies, common safety standards, and, where necessary, coordinated limits on development. Because some of those discussions could breach antitrust law, he also asks for a narrow waiver.

The diagnosis deserves to be taken seriously. If a laboratory slows down while its competitors continue, restraint can simply transfer advantage to the less cautious firm. Safety may therefore require coordination.

But coordination creates a second problem. An antitrust waiver can determine when firms may cooperate; it cannot by itself determine what risks the public should accept, which safeguards are adequate, or when development should resume. Those are not simply technical questions. They are judgments about how technological risk should be distributed across society.

If frontier AI has become too consequential to be governed by competition alone, then the answer cannot simply be to give its producers more freedom to govern collectively.

Amodei recognizes part of this problem. He regards regulation covering all frontier AI companies as the most effective approach, but also calls for voluntary cooperation in parallel. His explicit waiver request concerns certain safety conversations; protection for any resulting restrictions would raise a further question. Even a narrow permission can shape what follows. An interim agreement can establish which risks count, which tests demonstrate safety, and which firms are capable of satisfying them. By the time public regulation arrives, much of the governing architecture may already exist.

Coordination Is Not Authority

Safety alone does not settle who should exercise the authority a waiver would enable. Competition law has always done more than prevent collusion: It determines which forms of collective action the legal system permits. Labor unions, for example, can coordinate in ways that rival firms generally cannot. This is the allocation of coordination rights described by Professor Sanjukta Paul. An exemption for frontier AI firms would create a protected right for a small group of powerful companies to act collectively.

The U.S. Supreme Court confronted a version of this problem almost half a century ago. In National Society of Professional Engineers v. United States, engineers defended restrictions on competitive bidding because they feared that price competition would induce unsafe work. The concern was not frivolous, but the Court rejected the proposition that competitors could suppress competition simply because they regarded competition as dangerous. The public interest in safety did not automatically entitle an industry to determine how safety should be secured.

Frontier AI makes the problem harder because unilateral restraint may genuinely fail. Effective limits may require common triggers, monitoring, and credible commitments across firms, as AI researcher Nicholas Felstead has argued. Yet the more effective an agreement becomes, the more it begins to resemble the kind of horizontal restriction antitrust normally distrusts.

Congress is already considering S. 5105, the Collaboration on Adversarial Threats and Security Risks Act. Scholars Tim Schnabel and Dan Crane have supported this kind of narrowly drawn statutory protection for AI-security collaboration.

Suppose frontier laboratories agree not to train a class of models until specified evaluations are passed. An evaluator can determine whether a model passes the test. But someone must first decide which dangers deserve testing, how much evidence is enough, and what residual risk society should tolerate. Measurement can tell a regulator whether a threshold has been crossed, but it cannot tell society where the threshold should be.

When Antitrust Becomes Infrastructure

Silicon Valley was never simply the product of private markets. Federal research spending, defense procurement, and universities were integral to its development, as Professor Margaret O’Mara has shown. Competition law also shaped access to technology: the 1956 settlement with AT&T required broad patent licensing.

Decades later, digital platforms posed a different problem. Low consumer prices could coexist with growing control over the infrastructure on which other businesses depended, a tension at the center of Former FTC Chair Lina Khan’s 2017 critique of Amazon. The question was how antitrust should address that concentrated power.

Frontier AI now introduces an unusual reversal. The importance and potential danger of the technology become reasons for enlarging the collective rule-making capacity of the firms that dominate it.

A narrow waiver need not produce that outcome. The danger lies in the feedback between economic power and rule-making that Professor Luigi Zingales calls the “Medici vicious circle.” Firms might begin by exchanging information about security threats, then establish common evaluations and thresholds, and eventually coordinate pauses.

An agreement initially binds the firms that join it. If regulators subsequently adopt its tests, however, choices made by today’s leading laboratories can become conditions that tomorrow’s competitors must satisfy. And if regulators lack comparable expertise, they may increasingly organize their own decisions around standards the companies have already developed. What begins as an exception to antitrust can therefore become the infrastructure of later regulation.

Private Governance Moves Upstream

This would extend a form of private governance that technology companies already exercise. Large digital platforms have long written rules for speech, enforced them, and operated systems of appeal across spaces that millions of people use for public communication. Professor and journalist Kate Klonick called them the “new governors,” while Professor Julie Cohenhas shown how law itself helped constitute the informational markets and private powers that governments later struggled to regulate.

Anthropic makes this rule-making function unusually visible. Claude has a published constitution specifying values and behavioral priorities, with Anthropic occupying the highest position in the model’s hierarchy of principals, subject to external constraints including law. The document also expressly supports democratic institutions and checks on concentrated power. Publishing those choices is useful and arguably admirable.

For the people affected by these choices, the question is how to challenge them. Professor Gilad Abiri locates the democratic problem here: Explaining a model’s values leaves unresolved who is entitled to choose them. In Amodei’s proposal, the stakes extend beyond the rules for one company’s model. Outside evaluators would assess safety commitments, while competing companies could coordinate standards and limits affecting the technological frontier itself.

Those decisions should be open to challenge and revision through institutions whose authority does not depend on the companies’ consent. That requirement becomes especially pressing when private standards shape the opportunities available to firms outside the agreement.

Private standard-setting is indispensable to modern markets. Its effects nevertheless depend on who controls the process and how the resulting rules reach others. In Allied Tube v. Indian Head, producers of steel electrical conduit organized to prevent a competing plastic product from entering an influential industry code. Governments frequently relied on that code. The U.S. Supreme Court nevertheless denied the antitrust immunity the producers sought by presenting their conduct as an effort to influence government.

Antitrust scrutiny can address exclusion and collusion in such arrangements. Judging the adequacy of an AI safety standard would require a different kind of expertise: as Professor Herbert Hovenkamp explains, antitrust factfinders are poorly equipped to assess the substantive merits of a technical standard.

Amodei’s analogy between embedded AI evaluators and bank supervisors brings the question of authority into focus. Bank examiners do not possess authority because banks voluntarily invite independent experts inside and grant them access. Their authority comes from a legal regime that determines their mandate, rights of access, and consequences of noncompliance. The regulated firm does not decide whether supervision exists. Anthropic’s evaluators could have extraordinary technical access, but access is not authority, and transparency is not supervision.

Build Public Capability

Public authority can remain largely symbolic when governments lack the capacity to evaluate the systems they regulate. Elaborate rules may depend on weak science or evidence officials cannot independently interrogate, the danger legal scholar Matt Bartlett describes as “compliance theatre.”

Building that capacity does not require government to possess all the relevant knowledge itself. DARPA does not manufacture the technologies it supports, nor does it employ everyone capable of understanding them. Its program managers learn through dense networks connecting government, companies, and universities, the practice examined by Professor Erica Fuchs. They compare technological approaches, identify promising paths, and develop enough expertise to make independent judgments while relying heavily on knowledge generated outside the state.

Firms can build models, universities can research them, and independent organizations can evaluate them, while public institutions acquire enough technical competence to assess competing claims and decide what follows from the evidence. The NIST Center for AI Standards and Innovation offers a place to begin. It could be given more reliable funding, computing resources, and legally secured access to relevant systems, while retaining the ability to commission independent research and compare competing evaluation methods.

Experts can establish what the evidence supports, while legally authorized institutions decide what society should do with it. Under a mandate established by Congress, a public institution could then determine whether an identified danger justifies coordinated restrictions, whether the proposed safeguards are sufficient, and whether an antitrust exemption remains warranted, while the U.S. Justice Department continues to police the competitive consequences of the arrangement.

Any exemption should therefore also expire. Urgency may justify temporary protection for clearly defined cooperation, but continued protection should require an affirmative public finding that the danger and the restriction still warrant it. Decisions should state reasons, acknowledge uncertainty, and remain open to challenge and judicial review.

None of this guarantees good government. Public institutions can be captured, slow, and incompetent. But unlike an agreement among firms, they can be required to hear interests beyond those of the companies involved, explain decisions in public terms, and provide a legally recognized route through which those decisions can be revised. That is what Amodei’s own claim that “society must have a say” ultimately requires.

The danger is not merely conventional regulatory capture. Formal public authority can remain intact while the practical capacity to define and revise rules migrates elsewhere. Regulators can inherit standards they did not design, evidence they cannot reproduce, and technical dependencies that become increasingly expensive to unwind. Government can remain sovereign on paper while losing the capability required to exercise that sovereignty in practice.

Amodei is right that frontier AI may require moving beyond market incentives, but moving beyond the market does not require turning the firms that dominate it into a regulatory institution. This is the moment when the architecture of AI governance, and with it many of AI’s social consequences, is being set. How that architecture is built may ultimately matter more than any individual provision it contains.

Marco Mari is a PhD candidate in business and social law at Bocconi University and a research fellow at NYU Law School’s Program in Corporate Law and Policy and the MIT Industrial Performance Center.