Can AI companies team up on safety research and still beat each other in the race to the top?

This explores whether AI labs teaming up on common technical safety problems can fix the tendency to underspend on safety while they race each other. The short answer from the corpus is that collaboration helps only under specific conditions. Several sources also argue it isn't enough without outside enforcement. The corpus doesn't hold a direct study of joint industry safety research. It does have a lot on the incentives such projects would have to overcome.

The starting point is Can AI companies cooperate on safety despite competitive pressure?. Every company cuts safety corners because it fears rivals will cut theirs. The same note lists five conditions that make cooperation more likely: high trust, gains everyone shares, low exposure to being exploited, little competitive advantage from defecting, and shared losses if things go wrong. Shared technical challenges fit this frame well. Safety research that benefits everyone and gives no single lab an edge is exactly the kind of work companies can cooperate on. The harder question is whether the expensive kind of safety work, the kind that actually slows a product down, ever fits that description.

There is a less obvious catch. You might expect that more openness between labs would calm a race. A game-theory model in Does revealing rivals' capabilities reduce AI racing disaster risk? finds the opposite. When teams know exactly how capable their rivals are, the leader feels safe skipping precautions. When teams are uncertain, everyone stays cautious. So collaboration needs careful design: sharing safety techniques may help, but sharing capability information could make racing worse.

The technical challenges themselves are also harder to share than they look. Shared evaluations are the obvious place to collaborate, but Is evaluation awareness really one unified capability? shows that a model's ability to notice it is being tested breaks into separate parts that barely track each other. No single shared score reliably predicts how safe a model is once deployed. Likewise, Can individual components pass safety checks if the system still fails? shows that a system can pass every component-level check and still fail as a whole. And Do agents drift away from safety protocols during long interactions? finds that agents drift away from safety rules over long interactions, which one-off tests don't catch. An industry standard built on the wrong measurements could create a false sense of shared safety.

This is why much of the corpus looks beyond voluntary cooperation. Can companies alone manage the risks of AI systems? argues that companies can't police themselves and calls for binding government limits backed by hardware verification. Even OpenAI, in openai-argues-global-standards-may-pace-the-frontier-as-much-as-alignment-research, argues that international standards matter as much as alignment research itself, which is a quiet admission that technical work alone doesn't solve the coordination problem. Governments, though, are racing too. Can AI safety pacing work without government cooperation? describes Amodei's coordination proposal being rejected by the US and Chinese leaders within days. And Does slowing AI development actually prevent system failures? adds that even a well-coordinated slowdown only lowers risk; it never removes it. The corpus's overall view is that collaboration is a useful tool when the shared work gives no one an edge. It works best when paired with external enforcement and plans for responding to harm, and it can make things worse if the wrong information gets shared.