What would actually get rival governments like the US and China to cooperate on AI safety instead of racing ahead?

This explores what would get rival governments, especially the US and China, to coordinate on AI safety instead of racing. The corpus doesn't hand you a list of incentives, and it's worth saying that up front. What it does show is why goodwill alone hasn't worked, and which practical pieces would have to exist before any deal could hold. The clearest case is a failure. When Dario Amodei proposed pacing frontier development, both Trump and Xi rejected the idea within days. The lesson is that safety arguments lose when they run into national competition Can AI safety pacing work without government cooperation?. Any incentive that works will have to fit with how states compete, not ask them to set it aside.

One idea comes up again and again: cooperation depends on checking, not trusting. The Future of Life Institute argues that companies can't police themselves. It calls for government limits on AI systems that improve themselves, backed by hardware verification, meaning chip-level ways to confirm what others are actually running Can companies alone manage the risks of AI systems?. OpenAI frames the same problem as a collective action failure. If every lab and every country fears falling behind, no one slows down. On this view, shared international standards matter as much for pacing the frontier as alignment research does Can global standards pace frontier AI as much as alignment research?. Read together, these suggest the most useful 'incentive' may be infrastructure: tools that let a rival confirm you're keeping your side of the deal.

AI control research points the same way from an unexpected direction. Redwood Research argues that control is easier to check than alignment. Testing what a model *can* do is possible. Proving what it *intends* to do is not Can AI control work even if models are actively scheming?. That same logic applies between nations. Agreements built on capability tests and measurable thresholds are easier to verify than promises about good intentions. There's already a model for this: a frontier risk framework that scores models across seven capability areas against shared warning thresholds Where do frontier AI models actually pose the greatest risk today?. Its finding is a surprise. Today's models are already in the warning zone for persuasion and manipulation, while staying in the safe zone for cyberattacks and copying themselves. Persuasion threatens every government's domestic stability, so the shared danger may be closer than the usual 'rogue superintelligence' story implies.

One more point changes what a deal could realistically promise. Slowing development lowers risk but can't remove it, so agreements also need plans for responding when things go wrong, not just for preventing them Does slowing AI development actually prevent system failures?. That could make cooperation easier to sell. Countries that won't agree to freeze progress might still agree to shared incident response, the way rivals share warning systems for disasters. The gap to keep in mind: the corpus is strong on verification and on why pacing fails, but it has little on the economics of offering states something they actually want, such as market access, shared compute, or security guarantees, in return for restraint.