On September 12, 2026, Anthropic CEO Dario Amodei called for coordinated limits on the pace of frontier-AI development, with government support to address antitrust constraints. Sam Altman (OpenAI), Demis Hassabis (Google DeepMind), and Elon Musk (SpaceXAI) publicly supported the proposal’s direction. On September 15, OpenAI’s Chris Lehane said safety discussions with Anthropic and Google had been underway for weeks.
These public statements followed years of calls for coordination and more limited cooperation on safety. Since at least 2023, OpenAI, Anthropic, and Google DeepMind have exchanged ideas and coordinated on AI safety, even as they competed to develop more capable models. Examples include OpenAI leaders’ May 2023 proposal for leading developers to consider an agreed annual limit on frontier capability growth, the May 2023 joint call to make mitigating the risk of human extinction from AI a global priority, the July 2023 formation of the Frontier Model Forum to advance safety research and information sharing, and other AI executives’ public appeals for competing labs to work together to reduce AI risks.
What changed a few weeks ago was the nature of the desired collaboration. A prominent CEO publicly called for coordinated limits on competitive conduct, and rival CEOs endorsed the proposal within hours. The call for action recognized that the desired collaboration could violate federal antitrust law without government-granted immunity. Paying subscribers soon filed a class-action lawsuit against Anthropic, OpenAI, SpaceXAI, and Google in the Northern District of California. They alleged that the recent and earlier public statements amounted to an agreement to slow improvements in competing AI products, in violation of Section 1 of the Sherman Act.
These developments raise a pressing question: Can the frontier AI labs lawfully agree to slow innovation to address broad socially-oriented safety concerns without an antitrust exemption? The short answer is no. A bare agreement to slow innovation is clearly precluded by current antitrust law, regardless of whether it is meant to or does, in fact, benefit society. Frontier AI may be an area ripe for antitrust reform given the high stakes involved, but reform would take time and appears highly unlikely, at least in the foreseeable future.
Existing antitrust law nevertheless permits the frontier AI labs to engage in a range of collaborations that may have safety benefits. Given the grave risks the AI executives describe, we urge the AI labs to pursue permissible forms of collaboration vigorously. We outline how the labs might structure such collaborations.
A Bare Agreement to Slow Down Is Unlawful
Antitrust law is clear that a bare agreement among competitors to slow innovation does not become lawful merely because its stated purpose is to reduce risks to society, including human extinction. Such an agreement would fall squarely within Section 1 of the Sherman Act’s prohibition on unreasonable restraints of trade. The frontier AI labs are therefore correct in their assessment that they need government support to pursue what would otherwise be an unlawful arrangement.
Three principles govern the analysis. First, even a compelling safety objective cannot, by itself, excuse an unreasonable restraint of competition. Second, Section 1 requires an agreement; similar choices made independently do not suffice. Third, an agreement’s terms and competitive context determine whether the per se rule or the rule of reason applies.
Good Causes Are Not Enough. Section 1 of the Sherman Act prohibits agreements that unreasonably restrain competition. In assessing whether a restraint is unreasonable, the inquiry focuses on competitive harms to the parties’ trading partners and any countervailing competitive benefits. “Trading partners” are those who buy from or sell to the parties to the agreement—including customers, suppliers, and workers. The relevant harm must arise from impaired competition, such as higher prices, lower compensation, reduced quality, or diminished innovation. Even a grave injury counts as antitrust harm only if it stems from impaired competition. For example, a bridge collapse caused by negligent engineering may inflict catastrophic harm, but negligence alone does not make that harm anticompetitive. The analysis changes if the unsafe design resulted from an agreement among engineering firms to stop competing on quality.
Accordingly, social objectives unrelated to competition cannot, by themselves, immunize an otherwise unlawful restraint of trade. No matter how significant the risks of competition among frontier AI labs may be, those risks alone do not give rivals a license to suspend competition by agreement.
There Must Be an Agreement. A frontier AI lab may independently slow its pace of innovation to address safety concerns without violating the antitrust laws. Parallel conduct by several labs likewise does not, by itself, establish an agreement. Decisions made without an agreement may nevertheless be interdependent: Each lab observes and responds to its rivals’ conduct while anticipating their responses to its own behavior. As a result, a lab may be willing to slow development if it expects others to do the same yet accelerate if it expects them to press ahead. Such interdependence can produce similar conduct without an agreement. A parallel slowdown most likely would be unstable, as each lab retains a strong incentive to accelerate innovation and gain a market advantage.
An agreement may be proved by direct or circumstantial evidence. To support an inference of conspiracy, the evidence must tend to exclude the possibility that the defendants acted independently. At the pleading stage, the complaint must allege facts that, accepted as true, plausibly suggest an agreement. Allegations merely consistent with agreement do not suffice. Within this framework, an agreement may be inferred from an exchange of public statements. Competitors need not communicate privately to reach an understanding. Public announcements may support that inference when their content, timing, and surrounding circumstances indicate that competitors used them to propose and secure adherence to a common course of conduct. Public announcements and responsive conduct do not establish an agreement, however, when they reflect only independent decisions or conscious parallelism.
The Governing Standard of Review. Courts generally assess a Section 1 restraint under the per se rule or the rule of reason. Agreements subject to antitrust’s per se rule are deemed unlawful without any inquiry into their anticipated competitive harms and benefits. Agreements subject to the rule of reason are evaluated with reference to their competitive harms and procompetitive rationales.
Per se treatment ordinarily applies to naked agreements among competitors to fix prices, limit output, or divide markets, and to certain group boycotts or concerted refusals to deal. A stand-alone agreement among frontier AI labs to curtail development would most likely be treated as an output or quality restriction, and thus subject to the per se rule, particularly if it leaves customers with fewer or inferior products. But no controlling decision establishes that every agreement that, in effect, slows innovation falls within a per se category. The agreement’s terms and competitive context matter. Generalized safety concerns alone would not save a naked restraint. A limited restriction reasonably necessary to a genuine joint safety effort could instead receive rule-of-reason review and might still be unlawful.
Safety Collaboration May Justify Limited Restraints
While Section 1 prohibits a bare agreement among frontier AI labs to slow innovation, it does not categorically bar limited restraints tied to a genuine safety collaboration. Antitrust law permits rivals to form joint ventures, set safety and other standards, and exchange information in certain circumstances. The inquiry differs with the collaboration: It may turn on whether a restraint is reasonably necessary to a joint venture’s procompetitive purpose, whether a standard-setting process guards against anticompetitive bias, or whether the information exchanged could facilitate coordination. Across these settings, the concern is whether the challenged conduct serves the collaboration, or the collaboration serves as cover for it. Even if the challenged conduct genuinely serves the collaboration, its competitive effects still matter.
Three questions organize the analysis. First, is the restraint genuinely ancillary to a legitimate joint effort or a naked limit on competition? Second, how should an agreed delay in product releases be evaluated in its legal and economic context? Third, does the agreement restrict development itself, rather than merely delay release?
Naked vs. Ancillary Restraints. For horizontal restraints among actual or potential competitors, the distinction between “naked” and “ancillary” restraints can determine whether the per se rule applies. A naked restraint is an agreement to suppress competition that is not ancillary to a legitimate transaction or collaboration. Naked agreements among competitors to fix prices, limit output, or divide markets are unlawful per se. An ancillary restraint is a restriction on competition that is subordinate and collateral to a legitimate transaction or collaboration and reasonably necessary to achieve its procompetitive purpose. Ancillary restraints are evaluated under the rule of reason and may still be held unlawful. Merely placing a restraint in an otherwise lawful agreement does not make it ancillary.
An agreement among frontier AI labs to withhold newer models from the market could constitute a naked output restraint even if the labs invoke safety concerns. But a limited delay in deployment could qualify as ancillary if it is subordinate to a legitimate joint safety-testing program and reasonably necessary to obtain identifiable competitive benefits from the joint testing, such as improved model safety or reliability. The inquiry would focus on why obtaining those benefits reasonably requires the agreed delay, including its scope and duration.
Agreements to Delay Product Releases. An agreement among competitors to postpone introducing new or improved products can restrict output by withholding products that would otherwise be available. Whether it receives per se treatment depends on the agreement’s character and context. In Actavis (2013), for example, the Supreme Court required rule-of-reason review of reverse-payment patent settlements that postponed generic entry. Actavis, however, does not establish how to assess an agreed delay tied to safety collaboration.
For frontier AI labs, therefore, the analysis would depend on the agreement’s specific terms and the role of any resulting delay. A general commitment to postpone releases to address safety risks could amount to a naked restriction on competition. Invoking safety would not itself justify the restraint. An agreement to conduct joint evaluations of identified risks or develop safeguards could warrant different treatment, even if carrying out that work delays releases.
Agreements to Slow Innovation. An agreement to slow innovation restricts the development of new or improved products. By contrast, an agreement to delay product releases postpones their availability to consumers without necessarily restricting further development. Antitrust law recognizes that competition in R&D may form before products become commercially available, so the competitive inquiry need not be confined to current prices and quantities. An agreement that restricts independent research may therefore impair competition in developing new or improved products even before those products reach the market.
Designing a Safety Collaboration
If the frontier AI labs’ recent statements are taken at face value—and we see no reason not to credit those statements—artificial intelligence’s current trajectory poses a meaningful, existential risk to humanity. The frontier AI labs have proposed a collaborative innovation slowdown to manage that risk, but antitrust law prohibits such arrangements. However, that does not justify a complete absence of safety-related collaboration, especially since antitrust law affords the labs several options. Given the substantial stakes involved and the very low likelihood that the frontier AI labs’ desired antitrust reform manifests, the labs should quickly and zealously pursue the full extent of collaboration permitted by current antitrust law.
For frontier AI labs interested in joint safety work, we outline four core elements of possible collaborative frameworks: formal arrangements, governance and competitive safeguards, assessment of restrictions on independent conduct, and incentives and accountability. These elements could help structure projects aimed at producing identifiable competitive benefits for the labs’ trading partners. Whether a particular collaboration is lawful would depend on its terms, competitive context, and supporting evidence.
Formal Arrangement. The collaborating labs should document their collaboration in one or more agreements that define participants, projects, responsibilities, and resources. The agreements should distinguish joint work from independent activity and provide for revision as the work and evidence develop. For each project, they should identify the safety problem and intended output and explain, with supporting evidence, how the work could benefit the labs’ trading partners. Common evaluation methods, for example, could improve comparisons among models and reduce customers’ evaluation costs. The frontier labs also might consider structuring their collaboration as a joint research and development venture under the National Cooperative Research and Production Act of 1993 and filing the necessary notification, which could limit their antitrust risk and exposure.
Governance and Competitive Safeguards. The collaborating labs should establish procedures that assign technical decisions to qualified evaluators, require disclosure of relevant conflicts, and provide independent review of contested decisions. Participation and access criteria should guard against unnecessary disadvantages to smaller developers and exclusion of new rivals. Information exchanges should be limited to the project’s needs, with controls protecting competitively sensitive commercial and development plans. Review should examine the competitive effects of both the procedures and the resulting standards. Fair procedures alone would not ensure that the standards preserve competition.
Restrictions on Independent Conduct. The collaborating labs should identify and assess each proposed restriction on independent development or deployment. To justify a restriction as ancillary, explain why it is subordinate to the joint work and reasonably necessary to achieve its procompetitive purpose. Specify its scope, duration, and objective conditions for termination, and assess its competitive effects and any substantially less restrictive alternatives. A justification for delaying deployment would not itself justify restricting development. Even a restriction qualifying as ancillary would remain subject to rule-of-reason review.
Incentives and Accountability. The collaborating labs should explain why firms would participate and honor their commitments. Shared research costs and access to useful findings could encourage participation, but incentives to accelerate would persist. Specify who would monitor compliance and how unmet commitments would be addressed. Periodic review should compare actual results with expected benefits and identify needed revisions. These arrangements would also give the antitrust agencies or a court a concrete proposal to evaluate.
Voluntary collaboration among the leading AI labs—i.e., self-regulation—certainly may not meet the challenges posed by frontier AI. Moreover, the stakes are so high that society might not want to rely on the labs’ voluntary commitments alone. A broader regulatory framework with public oversight and enforceable safeguards is essential. Nonetheless, the labs should vigorously pursue lawful safety collaboration now, rather than wait for antitrust reform that may never arrive. Their collaboration must be part of society’s response, but society’s protection cannot depend on their willingness to collaborate.
Amelia Miazad is a professor at U.C. Davis School of Law, Barak Orbach is the Robert H. Mundheim Professor of Law & Business at the University of Arizona’s James E. Rogers College of Law, and Menesh Patel is a professor at U.C. Davis School of Law.