In June, Anthropic explicitly called for a globally coordinated AI pause mechanism. In July, over a thousand employees from the country’s leading AI companies published a public letter asking the US government to support an international effort to pace the frontier of AI development. Later, Sam Altman agreed that society “may have to pace the rate of AI development,” but warned that any coordinated slowdown would need to be done “in a way that doesn’t feel like regulatory capture for anyone, and also does not feel like collusion among the frontier labs.”
These public statements were a rare moment of harmony among fierce competitors. But Altman is also right to be cautious about avoiding collusion. At LASST, we’ve repeatedly fielded questions about whether and how antitrust law applies to coordinated safety activities among AI companies, so I thought I’d write up a plain-English summary of the basic legal landscape.
For background, I am a lawyer. I practiced at a firm for five years where I did a lot of antitrust law. This document is not meant as legal advice, but if you’re working in this area and want to collaborate, get in touch with us.
Legal background
There are two main antitrust laws: Section 1 and Section 2 of the Sherman Act. Section 1 applies to conduct involving two or more companies; Section 2 applies to conduct by one company.
For our purposes, the relevant law is Section 1, which prohibits agreements between companies to engage in unreasonable restraints of trade1. Cases tend to turn on two key questions: (1) was there an agreement to restrain trade, and (2) is that restraint of trade unreasonable.
Either an explicit or a tacit agreement can violate Section 1; what’s critical is that there is between the parties a “meeting of the minds,”2 “a conscious commitment to a common scheme.”3
Mere “conscious parallelism,” on the other hand—firms engaging in parallel business behavior with the awareness that their actions will inform how the others choose to behave—does not constitute a Section 1 violation.4
There are two frameworks for determining whether a restraint of trade is unreasonable: the “rule of reason” and the “per se rule.” The rule of reason is the presumptive framework. When a court applies the rule of reason, it weighs the totality of the circumstances, with the goal of “distinguish[ing] between restraints with an anticompetitive effect that are harmful to the consumer and restraints stimulating competition that are in the consumer’s best interest.”5 One significant consideration is whether the firms involved have market power.6
Some types of practices are so suspicious, however, that a court will forgo looking at the restraint’s effects and treat it as illegal per se. This is the “per se rule” framework. It applies only to restraints “that would always or almost always tend to restrict competition and decrease output.”7 When the per se rule applies, “[i]t makes no difference whether the motives of the participants are good or evil; whether the [restraint] is accomplished by express contract or by some more subtle means; whether the participants possess market control; whether the amount of interstate commerce affected is large or small; or whether the effect of the agreement is to raise or to decrease prices.”8 Restraints courts have deemed per se unlawful include agreements between competitors (known as “horizontal agreements”) to fix prices, divide up markets, and—critical here—restrict output.9
Agreement to pause/slow down
It would be unambiguously illegal for the leading AI companies to get together and agree to pause AI development, no matter how earnest and prosocial the motivations. This is an agreement to restrict output, a classic per se violation.
An agreement expressly to slow down AI development would likely meet the same fate. In practice, a slowdown would be implemented through practices tantamount to restricting output, e.g., no training models over X number of parameters, maybe no RSI. In National Macaroni Manufacturers v. FTC,10 leading macaroni makers (comprising 70% of the market) agreed to limit their products to a 50/50 durum wheat blend rather than 100% durum in response to durum wheat crop shortages. When the crisis passed, the macaroni makers went back to using 100% durum. The court found that their agreement was a per se Section 1 violation.
The Supreme Court made clear that agreements not to compete are illegal even if competitors have valid reasons for wanting to avoid competition. In National Society of Professional Engineers v. United States,11 engineers defended their profession’s ethical code against competitive bidding for projects along similar lines to the AI-safety argument for a pause: unrestrained price competition would push engineers to cut corners and produce unsafe work that endangers the public. The Court rejected their argument, holding that even “assum[ing] that competition is not entirely conducive to ethical behavior,” the law “does not support a defense based on the assumption that competition itself is unreasonable.”12
Agreeing on practices short of an outright pause
What if AI companies agreed to adopt certain practices (short of a pause or explicit slowdown) to mitigate AI risks? For example, they might agree to require third-party audits for new models, to publish system cards for every new model release, or not to train models on virological data. This gets complicated and very fact-specific. If an AI company ever gets the idea to do this, there will be many lawyers involved and the legality will depend heavily on the details. But I’ll outline some general principles for thinking about such agreements.
Some practices companies could agree to would have no antitrust implications. Antitrust laws concern themselves with practices that tend to suppress competition. Public disclosure of information generally does not present antitrust problems because it does not suppress competition—any competitor can access it. So an agreement among AI companies to publish system cards with every model release, or to publicize each company’s safety framework on its website, is not likely to present antitrust concerns. More on information-sharing is covered below, but AI companies sharing genuine safety and security relevant technical information (e.g. vulnerabilities, jailbreak techniques, child sexual abuse material/CSAM hashes) with each other is also generally not an antitrust problem.
Most practices probably would be assessed under the rule of reason. Only practices where the anticompetitive effects are extremely obvious get per se review. Many of the practices and standards AI safety people would like AI companies to agree to one day, short of outright pauses/slowdowns, would get rule-of-reason review.13
The rule of reason looks at the facts and considers whether procompetitive justifications outweigh the practice’s anticompetitive effects. Here are the steps to proving a Section 1 violation under the rule of reason: First, the government must show that the challenged restraint has a substantial anticompetitive effect in a relevant market—raised prices, reduced output, lower quality through reduced capabilities, etc. If the government succeeds in making that showing, the defendants have to offer a procompetitive justification. To prevail, the government must then show that the challenged restraint is not reasonably necessary to achieve the procompetitive justifications.14
A procompetitive justification may be hard to articulate. Notice how a procompetitive justification is needed. “This will meaningfully reduce existential risk” is a good reason to do something, but it wouldn’t be cognizable as a procompetitive justification because it doesn’t relate to competition. Suppose a bunch of AI companies all agreed that each will conduct third-party capabilities evals if the others do too. This arguably has anticompetitive effects because it slows the rollout of new models. A non-cognizable justification would be that third-party testing is important to ensure models don’t cause catastrophic harm to society. A cognizable justification would be something like, “enterprise customers are hesitant to adopt frontier models more widely in their systems because they don’t understand their capabilities well enough, so knowledge that a disinterested third party vetted the model will increase adoption.”15
But then you’d run into another problem: why is a mutual agreement “reasonably necessary to achieve” this procompetitive effect? Why can’t each AI company unilaterally decide whether to do third-party testing, and benefit from the competitive advantage of having third-party vetted products? It’s hard to answer this without implicitly admitting that the stable equilibrium of ruthless competition—which is what the antitrust laws are meant to encourage—may be rolling out new models as quickly as possible, testing be damned.
I think this is a problem many coordinated safety efforts between AI companies would face under Section 1. Good lawyers would probably figure out a way around it, though, in any given case.16 I’m just trying to show how antitrust law is intrinsically hostile to industry self-governance regimes premised on avoiding competition/racing dynamics.
How you enforce the industry agreement can independently give rise to liability. One complicating factor is that the means of enforcing an agreement to adopt a practice can independently constitute an antitrust violation—that is, an agreement to adopt even a competitively benign safety standard can turn into an antitrust violation if it is enforced in an anticompetitive way.
The biggest risk here is group boycotts, also known as concerted refusals to deal, which can be subject to per se antitrust liability. This happens when a group of competitors with market power effectively refuse to deal (i.e., buy or sell) with firms in order to competitively exclude/punish their rivals for a non-efficiency-related reason. For example, in Fashion Originators’ Guild, Inc. v. FTC,17 a group of fashion textile manufacturers comprising most of the market refused to sell to retailers that sold clothing with copied textile designs, aiming to deter IP infringement. The Supreme Court struck this down as a per se Section 1 violation, even though the purpose was to deter potentially illegal activity. So suppose OpenAI, Anthropic, and Google DeepMind decide that SpaceXAI is releasing models with insufficient safety controls, and agree among themselves not to make their models available to any deployer that also offers Grok models. That would likely be illegal.
Political activities, lobbying, other petitioning activities
Government-petitioning activity is shielded from antitrust liability under Section 1. This is called the Noerr-Pennington doctrine. OpenAI, Anthropic, and Google can’t strike a private agreement to pause AI development, but each could write a letter to Congress asking Congress to pass legislation ordering frontier AI companies to pause. They could also jointly write such a letter, because the doctrine shields coordinated government-petitioning activities. Under the same principle, the AI companies could jointly lobby the government (state or federal) to pass burdensome AI safety regulations that would have the effect of destroying new or disfavored AI companies. Intent would not matter here. “Joint efforts to influence public officials do not violate the antitrust laws even though intended to eliminate competition.”18
Information sharing
Sharing information to improve safety outcomes should be fine under the antitrust laws so long as the information is not competitively sensitive information, such as pricing, features and capabilities of future models, and customer lists. Sharing, e.g., the identities of bad actors or technical security flaws with competitors to improve systems industry-wide should be fine. The DOJ and FTC made this point in a joint guidance in 2014, explaining that antitrust law shouldn’t pose a barrier to legitimate cybersecurity information-sharing. More agency guidance on the permissibility of information-sharing may be forthcoming: DOJ and FTC recently solicited public comments for a forthcoming antitrust guidance on business collaborations.
AI safety and legal realism
I don’t believe that antitrust enforcement is an insurmountable barrier to AI companies coordinating more on safety. Antitrust law is a domain where courts are particularly flexible and pragmatic. There are even lots of activities that would have the practical effect of “slowing down” that a good lawyer could structure in a way that, at least, gets rule of reason review instead of per se review. Courts really hate to apply the per se rule, especially with new industries. I’ve described the capital-L “Law,” but it’s honestly probably more predictive to ask, “does it look like the AI companies are just bullshitting about why they’re doing this?” You’ll notice that in National Society of Professional Engineers, there are clear self-serving reasons aside from safety that engineers would want to delegitimize price-bidding as a practice. I think if a judge came away from the facts believing that AI companies really were coordinating out of pure public safety concerns, and the companies used the right words (e.g., NOT “coordinated slowdown”) and the right structures, the judge would grant the companies significant deference.
It’s also worth asking what the realistic downside is here. Criminal antitrust prosecutions are generally reserved for egregious cases of price-fixing. The exposure here is largely financial—costs of litigation, fines, damages—which are marginal compared to the opportunity cost of any coordinated measure that meaningfully slows down AI development. The companies would also, of course, be ordered to stop the illegal coordination. Plausibly, a court could also order more painful injunctions aimed at making sure similar conduct doesn’t recur, like placing an internal monitor at the companies to ensure they don’t engage in future coordination. Injunctive relief is flexible and would depend on the specific conduct challenged.
The text of the statute (passed in 1890) reads:
Every contract, combination in the form of trust or otherwise, or conspiracy, in restraint of trade or commerce among the several States, or with foreign nations, is declared to be illegal.
But courts subsequently grafted an “unreasonableness” requirement to prevent an overbroad interpretation that would ban all “restraints of trade,” since any commercial agreement, in a literal sense, restrains trade. See Chicago Board of Trade v. United States, 246 U.S. 231, 238 (1918) (“But the legality of an agreement or regulation cannot be determined by so simple a test, as whether it restrains competition. Every agreement concerning trade, every regulation of trade, restrains. . . . The true test of legality is whether the restraint imposed is such as merely regulates and perhaps thereby promotes competition or whether it is such as may suppress or even destroy competition.”).
Am. Tobacco Co. v. United States, 328 U.S. 781, 810 (1946).
Monsanto Co. v. Spray-Rite Serv. Corp., 465 U.S. 752, 768 (1984).
Bell Atl. Corp. v. Twombly, 550 U.S. 544, 554 (2007).
Leegin Creative Leather Prods. v. PSKS, Inc., 551 U.S. 877, 886 (2007).
Id.
Bus. Elecs. Corp. v. Sharp Elecs. Corp., 485 U.S. 717, 723 (1988).
United States v. McKesson & Robbins, Inc., 351 U.S. 305, 310 (1956).
Many cases use “price-fixing” to describe all agreements to fix prices, divide up markets, or restrict output because the direct economic effect of all such agreements is to fix quality-adjusted prices. I’m describing an agreement to pause as an output restriction here because it’s more intuitive.
345 F.2d 421, 427 (7th Cir. 1965).
435 U.S. 679 (1978).
Technically the Supreme Court applied the rule of reason to this case, even though it’s a clear price/output fixing restriction. Commentators have noted however that the Supreme Court did not undertake a full rule-of-reason analysis; the analysis resembled a per se analysis. This is often referred to as a “quick look” rule-of-reason analysis. The Supreme Court purported to do a rule-of-reason analysis instead because the court was wary of subjecting a professional society’s code of ethics to per se condemnation: “by their nature, professional services may differ significantly from other business services, and, accordingly, the nature of the competition in such services may vary. Ethical norms may serve to regulate and promote this competition, and thus fall within the Rule of Reason.” Nat’l Soc'y of Prof’l Eng’rs v. United States, 435 U.S. 679, 696 (1978).
See, e.g., Cal. Dental Ass’n v. F.T.C., 526 U.S. 756 (1999) (dental trade association’s ethics prohibition on discount advertising subject to rule-of-reason review); Wilk v. Am. Med. Ass’n, 895 F.2d 352 (7th Cir. 1990) (AMA’s ethical guidelines directing doctors not to refer patients to chiropractors subject to rule-of-reason review).
Ohio v. Am. Express Co., 585 U.S. 529, 541 (2018).
Cf. Cal. Dental Ass’n v. F.T.C., 526 U.S. at 771 (dentists’ agreement not to advertise discounts may be procompetitive because discounts tend to mislead patients on pricing, and greater pricing transparency is procompetitive).
For example, suppose instead of agreeing to do third-party evals generally, the AI companies agree to have the same third party evaluate each of their models using a standardized set of benchmarks. The companies could argue that this agreement is reasonably necessary because it allows their customers to compare directly between the capabilities of models in the market, encouraging usage.
312 U.S. 457, 460 (1941).
United Mine Workers v. Pennington, 381 U.S. 657, 670 (1965).