Could rival AI labs legally agree to slow down together, or would that count as illegal collusion?

This explores what would have to change in the law so that rival AI labs could agree to slow down together without that agreement counting as illegal collusion. The corpus names the legal gap clearly, but it doesn't offer a detailed legislative blueprint. The most direct material is the Congressional Research Service's analysis What antitrust risks do different AI safety collaborations face?. AI firms can already share cybersecurity threat data under existing exemptions. A pact to slow frontier development has no comparable protection, though, and to antitrust law it can look a lot like competitors agreeing to restrict output. CRS points to two routes: new legislation that creates an explicit safe harbor, or government waivers that approve specific arrangements case by case. The cybersecurity exemption is the closest template already in place. The open question is whether Congress would extend that kind of protection from sharing information to coordinating behavior.

The surprising thread across the collection is that legal permission may be the easy part. Karpf argues that a voluntary pacing plan with embedded evaluators, modeled loosely on bank supervisors, doesn't work unless the state can enforce it Can industry self-regulation slow AI without government enforcement?. Bank examiners matter because regulators can impose fines. By that logic, legislation that only makes coordination legal would produce a cartel that nobody can enforce. A law that actually works would probably need to make government a party to the arrangement, not just a bystander granting an exemption. Karpf also raises a conflict-of-interest point: the company proposing the pact may benefit from it, and that is exactly the situation antitrust law exists to examine.

Geopolitics adds a second layer. Amodei's pacing proposal Should AI capabilities growth be deliberately slowed to allow safety work? was rejected by both Washington and Beijing within days Can AI safety pacing work without government cooperation?. That suggests the real barrier is governments that see frontier AI as a strategic race. A domestic antitrust carve-out can't bind a foreign competitor, so any statute would face pressure to include exceptions that weaken it. Amodei himself argues that AI law should follow demonstrated risks, not anticipate them Should AI legislation wait for demonstrated risks to emerge?. That cuts against writing a broad pacing safe harbor before a clear harm justifies one.

Timing is also a problem. Legislative cycles take years while model releases take months Can regulation keep pace with AI's rapid evolution?. A fixed statutory exemption could be outdated before it takes effect. That points toward adjustable mechanisms, such as agency-granted waivers with sunset dates, over a single permanent carve-out. Finally, even a perfect pacing law would leave gaps. Slowing down lowers risk but doesn't eliminate failures Does slowing AI development actually prevent system failures?. It also doesn't settle who has the authority to stop a deployed system that is causing harm Can slowing AI development resolve who stops deployed systems?. Making pacing legal is one legislative problem, and deciding who can intervene in a deployed system is a separate one.