Over the next 1-5 years, I expect that the U.S.-China AGI relations will escalate because of three dynamics: (1) accelerated development of frontier AI, (2) China may increase its technical independence from U.S. export controls by producing domestic chips, (3) as AI systems become more autonomous they will become more opaque and harder to control.
The main risk in US-China relations is unintentional escalation given that both nations will deploy frontier systems in military and other applications as soon as they become available. Technical failures could be misattributed as adversarial action which could trigger disproportionate responses before the U.S. and China do an investigation. Current diplomatic channels don't have the required technical mechanisms to help distinguish an AI malfunction from a deliberate attack.
Intervention 1: Compute-Based Incident Reporting Protocol
There are no AI-specific notification protocols between the U.S. and China. Deep mistrust makes traditional self-reporting incident mechanisms politically impossible. I propose tying incident reporting to compute signatures which are observable metrics that could be tracked through enhanced export controls (including chip activity, power consumption, cluster size, etc.). This would allow the U.S. and China to verify potential AI failures independently without exposing their strategic capabilities.
Intervention 2: Joint Technical Standards
I recommend joint technical standards to identify and benchmark dangerous AI capabilities. Establish bilateral working groups that are hosted by neutral third parties (ISO or IEEE) to develop shared benchmarks for capabilities that both nations would consider high-risk. This can include autonomous cyber offense, biological design, strategic deception, and/or autonomous weapons integration.
Reporting Triggers: Tier 1 (24-hour notification): immediate safety threats, including AI-caused death or injury or models that demonstrate autonomous replication across distributed infrastructure. Tier 2 (48-hour notification): signal economic and/or infrastructure harm. This includes incidents during frontier-scale training runs (≥10²⁶ total FLOPs) that cause economic/infrastructure harm exceeding $100M.
To see the most success, this proposal should be presented through bilateral channels and framed as operational safety cooperation, rather than arms control with an 18-month pilot to begin immediately. Embed this in chip export strategy where compute access is conditioned on reporting participation.