Developing Safety Standards Is Not Inherently Anticompetitive
Four individual subscribers to Claude, ChatGPT, Grok, and Gemini have sued Anthropic, OpenAI, SpaceXAI, and Google, alleging that the companies agreed to slow the development of their competing AI products. The plaintiffs do not challenge the companies’ individual decisions about safety, testing, environmental impacts, or the pace of their own development. Instead, they argue that competitors agreeing on common safety standards can unlawfully limit the pace of innovation. But working together to develop safety standards does not, by itself, undermine competition.
Companies can cooperate on common approaches to safety while continuing to compete over the quality, capabilities, and pace of improvement of their products. The fact that a common safety standard may affect how quickly companies develop their products does not necessarily make the agreement anticompetitive. What matters is whether the cooperation restricts competition or instead helps companies compete safely.
Frontier AI labs are not exempt from the antitrust laws, nor should they be. Robust, neck-and-neck competition among these companies has helped make the United States the global leader in AI. Carveouts that allow companies to collude would undermine that competition, stifle innovation, and could ultimately weaken the United States’ position relative to China. But the opposite approach—using antitrust law to challenge legitimate cooperation on safety standards—could discourage competition and innovation. The goal should be to protect competition without treating every agreement among competitors as a threat to it.
The lawsuit is based on “We Must Pace the Frontier,” an essay published by Anthropic cofounder and CEO Dario Amodei on September 12. In the absence of federal regulation of AI development, Amodei called on frontier AI companies to embed independent evaluators in their labs and certify that increasingly capable models are commensurately well aligned to avert unintended negative outcomes like the Hugging Face incident and Australian government data breaches. Although he does not propose specific benchmarks or thresholds, he calls for “industry-wide coordination” on the capabilities of frontier AI systems based on their empirically validated safety. On the day the essay was published, the CEOs of SpaceXAI and OpenAI and the cofounder of Google DeepMind publicly expressed support for Amodei’s proposal.
Less than a week later, the plaintiffs filed their class-action antitrust lawsuit against the companies. The suit alleges that the AI firms illegally conspired to “reduce product quality relative to what unrestrained competition would produce.”
The lawsuit highlights that Amodei claimed that the federal government must issue a narrow antitrust waiver to permit the kinds of safety conversations he proposes. Plaintiffs are correct that Congress has not granted such an exemption, nor do the frontier AI labs possess one, but they don’t need one to collaborate on industry-wide safety standards. Indeed, the Standards Development Organization (SDO) Act of 2004 amended federal antitrust law to provide a rule-of-reason framework for covered standards development organizations, or SDOs, while they are engaged in standards development activities.
Amodei’s proposal should be understood as a call for an SDO. He explicitly calls on frontier AI labs “to establish common safety standards” and points to Google DeepMind chair Demis Hassabis’s framework for an AI standards body as one way to do so. Many industries already use similar organizations to develop voluntary standards, including telecommunications and networking. But companies do not necessarily need to create a formal SDO to develop common standards, and participation in standards-setting does not automatically violate the Sherman Act.
The courts have recognized this principle. In Consolidated Metal Products, Inc. v. American Petroleum Institute (1988), the Fifth Circuit ruled that the establishment and monitoring of trade standards is a legitimate function of trade associations and that standards-setting does not, by itself, establish an unlawful conspiracy. The court also warned that subjecting ordinary standards-setting to constant antitrust scrutiny could discourage the development of useful industry standards.
Moreover, standards-setting is not the only form of cooperation that Congress has recognized as potentially procompetitive. The National Cooperative Research and Production Act (NCRPA) provides a framework for companies to form qualifying joint ventures for research, development, and production. Rather than automatically treating covered joint-venture activity as illegal, the law requires courts to consider whether the cooperation actually harms competition, taking into account its benefits and competitive effects. Companies can also notify the Department of Justice and Federal Trade Commission of a qualifying venture, which can limit their potential antitrust damages to actual damages rather than the treble damages otherwise available.
The NCRPA is not an antitrust exemption: it does not protect agreements to fix prices, allocate markets, or impose other restraints outside the legitimate purposes of the joint venture. But its existence makes an important point. Federal antitrust law already recognizes that competitors sometimes need to cooperate to conduct research, develop technology, or produce something together without that cooperation being treated as inherently unlawful.
The same basic principle applies here. AI firms do not have to participate in Amodei’s proposal to offer their AI services to consumers, and companies that do not participate remain free to develop and sell their products. Meta CEO Mark Zuckerberg, for example, has publicly rejected Amodei’s proposal while Meta continues to develop its own AI products. The fact that some competitors support common safety standards therefore does not mean they have agreed to stop competing.
Amodei also called for “limiting the ingredients that go into frontier models, such as training compute.” If a company with market power actually withheld a critical input from rivals, that could raise a different set of antitrust concerns. But Amodei has no ability to control the supply of compute or data, and his call for government action to limit access to compute is itself protected speech. Moreover, Nvidia CEO Jensen Huang has publicly rejected Amodei’s pacing the frontier proposal. Given Nvidia’s dominant position in the market for the hardware required for AI model training and inference, neither Anthropic nor any other frontier AI company could unilaterally restrict U.S. developers’ access to compute.
The more difficult question is whether the companies’ public support for common safety standards could itself reduce competition. Even if those standards slowed the pace of frontier-model development, that alone would not establish an anticompetitive effect. The relevant question is whether the cooperation harms competition rather than simply changing the way companies compete.
The plaintiffs argue that future frontier AI models will be worse because of the alleged agreement than they would be under unrestrained competition. But that conclusion assumes that faster development necessarily produces better outcomes for consumers. If competition without common safety standards creates risks that could cause significant harm to consumers, then coordinating on safety could preserve competition by reducing the risk of a failure that prompts a much broader regulatory response or halts development. As Joseph Schumpeter observed, “there is no more paradox in this than in saying that motorcars are traveling faster than they otherwise would because they are provided with brakes.”
Robust competition has spurred recent advancements in AI and continues to propel global progress. Antitrust law should protect that competition. But protecting competition does not require treating every agreement among competitors as anticompetitive. Developing common safety standards does not, by itself, suppress competition, particularly when companies remain free to reject the standards and compete along this dimension of product differentiation.
The courts should therefore reject the plaintiffs’ theory that developing voluntary safety standards is itself an antitrust violation. If courts nevertheless conclude that existing antitrust law prevents frontier AI companies from cooperating on legitimate safety standards, Congress should make clear that it does not. Policymakers should not force companies to choose between competing vigorously and working together to prevent serious safety risks. The law should allow both.
