关键词: 人工智能安全, 中美关系, 政策文本分析, 治理合作, 全球人工智能治理

Abstract:

[Purpose/Significance] Artificial intelligence (AI) safety risks increasingly exhibit cross-border spillover effects, and the governance choices of China and the United States have important implications for global AI safety governance. This study aims to identify the policy issue structure and cross-national relationships in China-U.S. on this basis, AI safety governance and, on this basis, examine potential areas of cooperation under conditions of strategic competition. Compared with studies that mainly focus on broad governance models or individual policy differences, this study analyzes the relative position and cross-national relationship of specific AI safety governance issues within a unified policy framework and links the empirical comparison of policy issues with differentiated cooperation pathways. [Method/Process] Based on 50 AI governance policy documents issued by China and the United States between 2020 and 2026, 4 188 governance propositions were extracted through text screening and pre-processing. These propositions were treated as the basic units of analysis. To support cross-national semantic comparison, Chinese-language propositions were converted into English semantic representations, while policy texts originally written in English were retained. BGE-M3 was used to generate semantic embeddings, and BERTopic was applied for joint topic identification. To reduce the influence of differences in document length and proposition volume between the two countries, a country–document balanced topic discovery strategy was adopted, with repeated runs under different random seeds used to assess clustering consistency. The final model identified six governance topics. For cross-national comparison, document-weighted attention was calculated, together with MIX, BAL, and CAI. MIX measures the extent to which a topic contains policy content from both countries, BAL reflects the relative balance of policy attention between the two countries, and CAI captures the actual intensity of shared attention. [Results/Conclusions] The results identify six major governance topics. AI risk identification, assessment, and lifecycle governance, together with data security, privacy protection, and rights protection, receive relatively high bilateral attention. Standardization, testing, and evaluation show a certain degree of overlap in governance tools, but their policy weights remain asymmetric. By contrast, transparency and compliance responsibility, governance capacity building, and national-security-oriented AI infrastructure display more pronounced differences in policy priorities. These findings indicate that China-U.S. AI safety governance is characterized by the coexistence of shared risk concerns, limited technical overlap, and differentiated policy priorities. Based on the characteristics of different issues, cooperation may first focus on shared risks with clear cross-border effects and relatively low strategic sensitivity, including professional exchanges, risk evidence sharing, and comparative assessment. Standards, testing, and evaluation can serve as technical coordination interfaces because they offer relatively comparable governance tools and procedures. Issues closely associated with frontier capabilities, critical infrastructure, and national security are more suitable for limited communication, risk-control arrangements, and crisis-management mechanisms. The depth of cooperation should therefore be determined by cross-border spillover effects, the coordinability of governance tools, and strategic sensitivity, with government communication channels, scientific and expert networks, standards and technical organizations, multilateral mechanisms, and Track II networks assuming corresponding functions. This study is limited to formal central- and federal-level policy documents, so the findings reflect policy issue structures and potential cooperation space rather than implementation outcomes. Future research may incorporate bilateral dialogues, international organization documents, expert networks, and concrete cooperation cases to examine how shared policy attention can be translated into sustained governance cooperation.

Key words: AI safety, China-U.S. relations, policy text analysis, governance cooperation, global AI governance