AI Pacing Agreements: Legal Issues And Potential Strategies To Manage Them

AI pacing agreements—coordinated commitments among frontier developers to slow capability rollouts until safety research and oversight catch up—could help reduce catastrophic risks. They also raise substantial legal questions, especially under U.S. antitrust law, and implicate a range of other regulatory, contractual and governance issues. This article outlines the key legal challenges and practical strategies for companies evaluating such arrangements.

Why Pacing, And What It Might Look Like

Industry leaders have called for deliberate pacing as model capabilities accelerate and safety, alignment and governance struggle to keep up. A commonly discussed framework includes three escalating steps:

• Step 1: Embedded evaluators. Independent third-party assessors with ongoing access to internal safety processes to verify compliance, monitor incidents and provide external validation—akin to supervisory models used in other regulated sectors.
• Step 2: Democratic coordination. Alignment on capability checkpoints and minimum safety certifications before further training, deployment or use in recursive self-improvement. Proposals also contemplate coordination on compute, training-run structure and use of AI to build AI—each posing distinct antitrust concerns.
• Step 3: Global coordination. International standards with verified compliance and inspection mechanisms, balancing safety with national security and competitiveness.

Antitrust: The Central Legal Challenge

Sherman Act Section 1 prohibits unreasonable restraints of trade. Pacing agreements among direct competitors can be characterized as output restrictions—agreements to limit the quantity, quality or speed of production—long viewed as paradigmatic restraints. That framing risks per se illegality if the agreement appears to be a “naked” restriction unconnected to a legitimate collaboration.

The safer pathway is to structure any restraint as ancillary to a bona fide joint safety venture that creates pro-competitive value (for example, shared evaluation protocols, red-teaming infrastructure or a recognized certification “product”). Under the ancillary restraints doctrine, reasonably necessary and subordinate limits tied to that collaboration are analyzed under the rule of reason rather than condemned automatically. Success turns on careful design: demonstrate why limited pacing is needed to make the safety collaboration effective, narrowly tailor any restrictions and preserve independent decision-making where feasible.

NCRPA Safe Harbor—Helpful, But Limited

The National Cooperative Research and Production Act (NCRPA) encourages joint R&D by assuring rule-of-reason treatment and limiting damages for qualifying ventures. However, it excludes agreements that restrict production or distribution. Because pacing involves limiting capability development and deployment, those elements likely fall outside the NCRPA. A layered approach can help:

• Register the joint safety research venture (evaluation methodologies, shared testbeds, red teaming) under the NCRPA.
• Address any pacing elements separately as ancillary restraints, supported by narrow tailoring and governance safeguards.

Targeted legislation creating an AI safety coordination safe harbor—akin to tailored information-sharing protections adopted in other risk domains—would provide firmer footing.

Regulatory Tools, State Law And Private Litigation

DOJ Business Review Procedure. Parties can seek a business review letter from the Antitrust Division outlining current enforcement intent regarding a proposed framework. Though not binding immunity, a favorable letter offers meaningful comfort and can deter challenges while legislation evolves.

Shifting guidance. Federal collaboration guidance has been in flux, heightening uncertainty. Well-structured proposals emphasizing consumer benefits, safety outputs and limited, necessary restraints are more likely to receive a measured response.

State antitrust law. States may apply different or stricter standards; active enforcement by state attorneys general increases risk. Multistate operations require jurisdiction-by-jurisdiction analysis.

Private suits. Treble-damages litigation by excluded rivals, customers or startups is a material risk. Plaintiffs must show antitrust injury, but the defense burden and discovery exposure are significant—another reason to document pro-competitive justifications and narrow tailoring.

Beyond Antitrust: Other Legal Exposures

• IP and trade secrets. Safety data can reveal sensitive information (architecture, scaling trajectory, training methods). Use clean teams, strict confidentiality, access controls and express IP carve-outs; ensure “reasonable measures” to preserve trade secret status.
• Liability and indemnification. Allocate risks if harms occur despite passing shared standards, or if pacing allegedly delays beneficial releases. Consider reciprocal indemnities, insurance requirements and clear standard-setting disclaimers.
• Export controls and national security. Deemed-export rules and controls on model weights, code and dual-use tech require careful classification, nationality-based access controls and recordkeeping—especially if embedded evaluators include foreign persons.
• International regulation. Align with the EU AI Act and other regimes; competition law under Article 101 TFEU differs from U.S. law. Multinational arrangements need tailored compliance.
• Governance and fiduciary duties. Boards should document informed, good-faith decisions balancing safety, competitiveness and oversight obligations; robust records help under business judgment and oversight doctrines.
• Government engagement. Joint advocacy for legislation or regulation benefits from Noerr-Pennington protections; keep advocacy distinct from any substantive pacing commitments.
• State frontier-AI requirements. Emerging state laws on audits, incident reporting and third-party assessments may overlap with or constrain voluntary pacing frameworks.

Practical Guidance For Structuring Pacing Agreements

• Center the arrangement on a genuine joint safety venture that produces verifiable outputs (evaluation protocols, certifications), with any pacing limits narrowly tailored and time-bound.
• Register eligible research elements under the NCRPA; separate any output-affecting measures and justify them as reasonably necessary ancillary restraints.
• Seek a DOJ business review letter with comprehensive detail on scope, governance, information flows, and how competition elsewhere remains robust.
• Use independent third-party evaluators; implement clean teams, data minimization and role-based access for sensitive information.
• Preserve independent decision-making on competitively sensitive variables where possible; avoid sharing prices, customer allocations or unrelated strategy.
• Adopt clear capability checkpoints tied to objective tests and pre-defined safety thresholds; include off-ramps, sunset dates and periodic reauthorization.
• Implement antitrust compliance training, audit trails and a governance committee (including external advisors) to oversee adherence and revisions.
• Address IP, confidentiality, export controls and data residency in the contract; classify data and restrict access by nationality where required.
• Allocate liability via indemnities, insurance and disclaimer language; define processes for incident response and standard updates.
• Coordinate with regulators and legislators transparently; separate protected joint petitioning from operational coordination.

Conclusion

Pacing agreements present real antitrust exposure if structured as naked output restraints. With careful design—anchoring limits to a bona fide safety collaboration, separating NCRPA-covered research from any pacing components, obtaining DOJ business review guidance, and deploying robust safeguards—companies can reduce risk while advancing meaningful safety goals. Continued government engagement and targeted legislation would materially improve legal certainty. Organizations considering participation should engage counsel early, document pro-competitive and safety rationales, and build adaptive governance that can evolve with fast-moving legal and technological developments.