Racing dynamics

When companies and countries compete to build the most powerful AI first, each one is pushed to cut corners on safety.

A race is when being first matters so much that everyone speeds up. The AI Safety Atlas describes AI development as increasingly like a “winner-take-all” contest, where “competitive pressures systematically undermine safety investments when speed provides decisive advantages”. The Center for AI Safety counts the “AI race” as one of four main sources of catastrophic AI risk, because competition can “compel actors to deploy unsafe AIs or cede control to AIs”.

It’s a collective-action problem: each company may prefer to go more carefully, but fears that slowing down only hands the lead to a less careful rival. The same logic applies between countries. The Atlas compares it with medicines, where regulation and liability make a race to the bottom on safety rare, and notes that AI has so far had little of either.

That’s why many proposed fixes are about coordination rather than technology: shared rules like the EU AI Act, independent testing before release, international agreements, or a pause. In March 2023 the Future of Life Institute’s open letter asked AI labs to pause training systems more powerful than GPT-4 for at least six months. No pause happened.