Antitrust Law May Block AI Safety Pauses Among Competitors

The Legal Barrier to Coordinated AI Safety Measures

Anthropic recently published an essay warning that AI systems may soon become capable of autonomously designing their own successors—a threshold called "recursive self-improvement." The company stated it would be willing to slow or temporarily pause frontier AI development if peer companies did the same in a verifiable manner. But this safety-minded proposal runs headlong into a fundamental problem: U.S. antitrust law generally prohibits competitors from agreeing to restrict their output.

The tension highlights a critical gap in AI governance. As Nicholas Felstead writes in Lawfare, which first reported these details, the very features that would make a coordinated pause effective—binding commitments, verification mechanisms, and consequences for defection—are the same elements that establish an illegal agreement among competitors under Section 1 of the Sherman Act.

Why Unilateral Pauses Won't Work

Anthropic's position is that a single lab pausing development would accomplish little beyond changing which company leads the race. The company notes that letting "the least cautious actors catch up technologically ... could leave everyone less safe." This race dynamic creates a collective action problem: each lab believes it is more careful than its rivals, so each concludes it should continue development.

METR, a leading AI evaluation institute, has found that the length of software tasks AI systems can complete autonomously doubles roughly every four months. Anthropic's internal data supports this trajectory. The company acknowledges that alignment research and governance are not keeping pace with capability advancement, and that "rare occurrences of misalignment present in today's models could compound as the models build their successors."

For a pause to be meaningful, Anthropic argues, it must involve "multiple well-resourced labs at or near the frontier, in multiple countries, agreeing to stop under the same conditions" with mutual verification.

The Per Se Problem

Agreements between competitors to limit production or development are typically condemned as illegal per se under antitrust law—meaning courts don't examine justifications or weigh benefits. As the Supreme Court stated in NCAA v. Board of Regents, "Horizontal price-fixing and output limitation are ordinarily condemned as a matter of law" because "the probability that these practices are anticompetitive is so high."

A coordinated pause is, in form, an agreement among competitors to stop developing and releasing products. Even if analyzed under the more forgiving "rule of reason" standard, safety justifications face an uphill battle. In National Society of Professional Engineers v. United States, the Supreme Court rejected an engineering association's argument that its ban on competitive bidding protected public safety, calling it "a frontal assault on the basic policy of the Sherman Act."

The immediate risk facing AI labs is not just whether they might ultimately prevail in court, but whether they want to face years of litigation, treble damages, and enforcement actions from state attorneys general.

Why It Matters

The antitrust barrier to coordinated safety pauses reveals a fundamental misalignment between existing competition law and the unique risks posed by frontier AI development. Unlike typical markets where competition reliably produces better outcomes, AI capabilities may be advancing faster than the industry's ability to ensure safe deployment. If the most safety-conscious labs cannot legally coordinate to slow development when facing catastrophic risks, the regulatory framework may be inadvertently accelerating a race to the bottom.

Potential Solutions

Felstead suggests several approaches. The Department of Justice and Federal Trade Commission have opened a public inquiry on competitor collaborations, and several submissions have raised AI safety coordination. New guidelines could clarify when pausing arrangements might be analyzed under the rule of reason rather than condemned outright.

The cleanest solution would be congressional action: a statute requiring frontier developers to halt development when specified capability thresholds are breached. Compliance with a legal mandate is not a conspiracy among competitors. However, such legislation faces difficult design questions and appears unlikely in the near term.

Felstead emphasizes this is not a case for exempting AI labs from antitrust law generally. The industry has real concentration concerns, and enforcement will remain a priority. But a narrow, carefully constructed mechanism for safety-based coordination may be necessary to prevent the race dynamic from overriding prudent risk management.

These details were first reported by Nicholas Felstead in Lawfare.