Experts urge U.S.-China cooperation to curb AI risks without slowing innovation
Two technology policy specialists are calling for the United States and China to collaborate on artificial intelligence safety, warning that increasingly powerful systems pose cross-border risks that neither country can manage alone. In a new commentary, Christina Knight, a J.D.
and M.B.A. candidate at Harvard University who previously led Scale AI’s Security and Policy Research Lab and served as a senior policy adviser at the U.S.
Center for AI Standards and Innovation, and Scott Singer, a fellow in the Technology and International Affairs Program at the Carnegie Endowment for International Peace, argue that AI models can be misused to engineer dangerous pathogens, launch autonomous cyberattacks, or spread realistic deepfakes regardless of the user’s location.
They write that neither Washington nor Beijing benefits from an AI race in which a model from either country could cause catastrophic harm anywhere. The authors highlight what they describe as particularly acute vulnerabilities in some Chinese models. DeepSeek’s open-source large language model, R1-0528, lacks many safeguards built into U.S.
systems and accepts malicious instructions 12 times more often than leading U.S. models do, according to U.S. government research cited by the authors. They also say standard jailbreaking methods elicit harmful responses 94 percent of the time from such systems, compared with 8 percent for comparable American models.
Those risks, they warn, can multiply when a model powers autonomous agents capable of browsing the web and accessing databases without human oversight, such as OpenClaw. Knight and Singer draw on Cold War–era precedents to argue that open communication channels are essential when great powers develop high-risk technologies.
They note that U.S. scientists shared information with the Soviet Union to help prevent unauthorized nuclear use, suggesting that rivals can still cooperate on guardrails while competing elsewhere. Their proposal centers on joint risk mitigation rather than curbing progress.
They argue that a prudent U.S. approach does not mean slowing innovation; instead, it involves working with Beijing to agree on safety research priorities, coordinate testing for vulnerabilities, implement safeguards, and jointly establish best practices to contain global risks.
They add that China should invest in the technical capacity needed to make such engagement worthwhile. The authors also call for systematic assessments of frontier AI—akin to clinical trials for drugs or crash tests for cars—to identify dangers before and during deployment.
But they caution that AI differs from traditional technologies because systems are general-purpose, evolve after release, can be repurposed in unanticipated ways, and spread globally at unprecedented speed. Testing before launch alone, they argue, is not sufficient.
Knight and Singer contend that Washington and Beijing can compete fiercely on AI while still mitigating extreme dangers if they focus cooperation on how to identify risks rather than on sharing the specifics of what they find. With the right approach, they conclude, making AI safer is both necessary and feasible.