What antitrust risks do different AI safety collaborations face?

The Congressional Research Service (CRS), the non-partisan research arm of the U.S. Congress, states that "[d]ifferent types of safety collaboration among AI developers would raise different levels of antitrust risk under current law." It finds that "the sharing of information regarding cybersecurity incidents and threats—standing alone—would likely involve low levels of antitrust risk," protected in part by the Cybersecurity Information Sharing Act of 2015 (CISA)'s antitrust exemption and eligible for rule-of-reason review as a research-and-development joint venture under the National Cooperative Research and Production Act of 1993 (NCRPA). By contrast, Anthropic chief executive Dario Amodei's September 12, 2026 call to "pace the frontier" — coordinating AI companies on common safety standards and "limits on the rate of unchecked AI progress" — is the broader kind of coordination CRS does not extend the same protection to; Amodei himself acknowledged that "[s]ome forms of coordination that would be impactful for pacing are legally challenging, and will require government support," potentially including "waivers of antitrust restrictions."

CRS's reasoning rests on specific legal mechanics. The DOJ and FTC's 2000 Antitrust Guidelines for Collaborations Among Competitors treat most R&D joint ventures as procompetitive and subject to rule-of-reason analysis rather than per se illegality. The NCRPA lets qualifying R&D collaborations limit damages exposure to "actual damages plus costs and attorneys' fees," rather than Sherman Act treble damages, by filing a notification with the DOJ and FTC. CISA's cybersecurity exemption — "set to expire on December 11, 2026, absent legislative extension" — specifically covers threat-indicator sharing while excluding price-fixing, market allocation, or exchange of competitively sensitive information. None of these mechanisms were written with "pace the frontier" coordination in mind, which is why CRS treats it as comparatively uncertain territory that may need new legislative action rather than an existing exemption.

Amodei's pacing proposal sits alongside his separate argument, in Should AI legislation wait for demonstrated risks to emerge?, that legislation should follow demonstrated risk — this CRS sidebar shows what happens when that legislative follow-through has not yet arrived: companies attempting coordination without a clear antitrust safe harbor, and customers filing a Sherman Act class action in response within days. The legal risk CRS describes is the obstacle side of the problem named in Can AI companies cooperate on safety despite competitive pressure?, whose five tipping factors describe why firms might want to cooperate on safety; antitrust law determines whether they are legally free to act on that want. It also bears on Can three-tier AI oversight actually prevent deployed system harms?, which assigns companies pre-deployment testing duties without addressing whether coordinating to meet them would expose U.S. firms to antitrust liability, and on Does revealing rivals' capabilities reduce AI racing disaster risk?, whose formal race dynamics motivate exactly the kind of inter-firm coordination this sidebar finds legally unresolved.

The excerpt does not say whether the September 18, 2026 Sherman Act class action will succeed, or whether Congress will act before CISA's cybersecurity exemption lapses on December 11, 2026; it reports a legal landscape and a live dispute, not an outcome. It also does not evaluate whether antitrust waivers are good policy — only that current law offers no clear exemption for development-pacing coordination comparable to the one it offers for cybersecurity-threat sharing. The implication is narrow but concrete: voluntary industry coordination of the kind Amodei proposed likely stays legally contested, or dependent on new legislation, rather than proceeding under existing antitrust guidance.