U.S. and Chinese Experts Find Common Ground on Guardrails for AI in Biology

NTI | bio and the Center for International Security and Strategy (CISS) at Tsinghua University convened a virtual dialogue on risks at the convergence of artificial intelligence and the life sciences (AIxBio). U.S. and Chinese experts from academia, industry, and civil society agreed that AIxBio capabilities are advancing faster than the efforts to govern them. Participants noted that the agreed U.S.-China intergovernmental dialogue on AI offers an opportunity to narrow these gaps and help both countries realize the benefits of fast-advancing AIxBio capabilities while reducing the risk of misuse.

The two-day, virtual event in July revealed opportunities to advance technical cooperation between U.S. and Chinese experts on AIxBio risk reduction, including joint development of a risk taxonomy for AIxBio capabilities and discussion of approaches to access governance. Both actions would strengthen safeguards against misuse while advancing responsible research and innovation. Xiao Qian, Deputy Director of CISS and Vice Dean of the Institute for AI International Governance at Tsinghua University, and Jake Jordan, Vice President of Global Biological Policy and Programs (NTI | bio), co-chaired the dialogue, emphasizing the importance of U.S.-China cooperation.

Experts examined the AIxBio risk landscape across four areas:

1. Risk assessment and horizon scanning
2. Technical guardrails for biological AI models
3. Access frameworks for tools and data
4. Emergency response and crisis communication

U.S. and Chinese experts converged on several points. Participants agreed that effective risk reduction requires a strong evidence base for understanding emerging capabilities and risks. They also agreed that guardrails developed for large language models, such as secure model weights, use-restriction classifiers, and managed access, should be matched by comparable safeguards for biological AI models. In addition, experts identified managed access to biological AI models and the underlying data as an area requiring immediate attention from both countries, including shared standards for validating legitimate users and ensuring ongoing oversight after access is granted.

Participants recognized the importance of creating clear communication channels as part of preparedness for AIxBio-related risks. They discussed the value of exploring appropriate, duly authorized communication mechanisms among competent authorities and relevant stakeholders, consistent with respective laws. This would help manage the risk of misunderstandings and enable authorities to respond more effectively in a crisis.

NTI | bio and CISS plan to co-author two papers on AIxBio risks and possible risk-reduction approaches. Drawing on collaborative analysis by U.S. and Chinese experts, the two papers aim to contribute constructive input to AIxBio risk reduction and biosecurity governance. The papers will present the authors’ analysis and may contribute to future U.S.-China policy dialogues.