Why It Matters
Competing developers of frontier artificial intelligence models have no settled path to jointly slow work for safety without inviting unlawful-restraint claims, a new Congressional Research Service report finds.
On September 12, Anthropic chief executive Dario Amodei urged peers to pace advanced systems to curb loss of control and misuse for cyberattacks and bioterrorism, proposing shared standards and limits on unchecked progress that could need government backing including waivers.
Other lab leaders endorsed pacing; House Minority Leader Hakeem Jeffries sought decisive congressional action; and critics said current law already permits much safety work and waivers could harm competition. Separate commentary cited rivalry with the People's Republic of China and framed a slowdown as pretext to lower capital outlays.
On September 18, customers of several leading companies filed a putative class action alleging an agreement to slow innovation.
The Big Picture
Section 1 of the Sherman Antitrust Act, as codified, bars every contract in restraint of trade but is read to ban only unreasonable harms, with price fixing, market division, and output limits condemned without inquiry into effects.
Most other restraints face fact-specific inquiry where plaintiffs must first show substantial harm such as higher prices, reduced output, or diminished innovation, defendants must then show a competition-tied justification, and plaintiffs may still win with a less restrictive alternative, with some courts adding final balancing.
In National Society of Professional Engineers v. United States, the Supreme Court rejected a ban on competitive bidding defended as protecting public safety, explaining the statute reflects a judgment that competition produces better goods and services.
An intermediate quick-look standard lets plaintiffs meet their initial burden without detailed proof of market power for conduct resembling per se violations, while still allowing justifications.
Research and development joint ventures were typically subject to the rule of reason under 2000 agency guidance withdrawn in December 2024 as no longer providing reliable guidance, though case law still supports that approach, and the joint venture label cannot save an otherwise inherently illegal agreement.
Information exchanges alone generally face the same full inquiry, with price, output, costs, strategic planning, and current or future plans drawing more concern than historical material, and 2014 guidance said cyber threat sharing appeared unlikely to raise price or output concerns.
Sharing cybersecurity incidents standing alone presents likely low risk generally analyzed under the rule of reason consistent with 2014 agency policy, while sharing alignment methods presents low to moderate risk and joint testing of prerelease models may pose heightened risk, with separate protection under the Cybersecurity Information Sharing Act of 2015 ending December 11 absent legislative extension.
Third-party evaluators could reduce that exchange problem, unilateral use needs no agreement to trigger liability, and pacts among firms to use evaluators turn on details.
A pact to withhold models lacking third-party certification under common criteria raises nontrivial risk, defensible as private standard setting where standards rest on objective expertise through unbiased procedures, though concerted enforcement faces stricter review.
A July 2025 enforcement statement distinguished arrangements leaving firms free to sell nonconforming products from those forbidding them, and a 2025 inquiry into truck makers honoring California emissions rules closed in August 2025 after pledges to act independently.
A coordinated halt above capability thresholds carries the greatest exposure as a horizontal curb on quality and research competition, potentially treated as ancillary if tied to broader evaluation or condemned without inquiry if standalone.
The National Cooperative Research and Production Act of 1993, as enacted, directs reasonableness review for covered ventures and limits filers who notify officials to actual damages plus costs and fees rather than treble damages.
In the 119th Congress, the Collaboration on Adversarial Threats and Security Risks Act would excuse good-faith exchange on covered security risks and coordinated delay upon written notice to the head of the Antitrust Division, structured as an affirmative defense proved by preponderance.
Narrower choices include an information-only excuse or renewal of cyber sharing cover, while H.R. 9925 would build Commerce-administered oversight and tort principles may under-deter catastrophes that leave a developer unable to pay.
The Bottom Line
For the current administration, the near-term question is less whether safety cooperation has value than which institutional channel absorbs residual legal doubt.
With operative research guidance withdrawn and enforcement signals skeptical of feature restrictions, executive actors face pressure to clarify assessment of certification and evaluation pacts without inviting collusion claims.
Lawmakers confront an expiring cybersecurity sharing protection, competing designs for broad exemptions versus narrow exemptions versus upfront regulation, and a pending court test of whether safety justifications survive precedent.
Access the Legis1 platform for comprehensive political news, data, and insights
Access the Legis1 platform for comprehensive political news, data, and insights.
Spot something wrong? Report an issue with this article