U.S.-China AI Agreements Need Verification, Not Trust

The upcoming Trump-Xi summit is unlikely to yield an ambitious U.S.-China AI safety agreement due to deep mutual distrust, despite calls from AI lab leaders for urgent coordination. However, new verification technologies could make future agreements feasible without requiring political trust. The U.S. and China have established a bilateral AI dialogue, but neither is prepared to accept substantive constraints on AI development. Verification mechanisms, such as satellite imagery, cryptographic proofs, and network traffic shaping, could monitor compliance with agreements limiting compute use or training runs, though such infrastructure requires years of development. Arms control precedents, like Cold War reconnaissance satellites, suggest verification can enable cooperation between rivals. The U.S. government should fund verification R&D now, as the field has only about 50 engineers worldwide, to prepare for future agreements.