A group of AI researchers wants to end the AI race by law. Their proposal would make it illegal to train AI systems beyond a fixed amount of training. A single multinational institution would enforce that legal ceiling and oversee any work above it. The aim is to stop a contest in which whoever spends most on computing power pulls ahead in what AI can do.
Another proposal, for an international agreement, puts numbers on such a cap. It measures training by the number of calculations a computer performs. Any training run above a trillion trillion calculations would be banned, while runs above a hundredth of that would need approval and monitoring. How much training a company runs would stop being its own competitive choice and become a governed activity.
The researchers’ logic rests on incentives. In a race, teams are rewarded for arriving first, so safety work that slows them down comes under pressure. A formal model of AI races shows how competition can lead developers to cut safety precautions to gain speed. Self-regulation is least reliable when competitors fear being left behind.
That is why the first proposal puts a safety gate before a system is built, instead of a fix after launch. Scaling up would require permission, and permission would depend on evidence that the risk stays below an acceptable level. Evidence would become a condition for training the most advanced models, which would slow their time to market.
Enforcement is the hard part. The agreement would check the limits through computing power itself, by tracking AI chips and verifying how they are used. Its rules would then reach supply chains, computing providers and the companies that run large banks of chips, as well as what the models produce. That is what makes a cap on computing power easier to enforce than a promise.
If this becomes policy, companies could not buy a leap in capability through larger training. Advantage would shift to how well they use the AI that already exists, through lower cost per task, fewer errors and faster cycle times. Privileged access to the biggest training runs would become more valuable, and operational excellence more valuable still, because most firms will work below any cap.
