关键词: AI 虚拟馆员, 大语言模型, 人机协同, 高校图书馆, 在线参考咨询, 检索增强生成（RAG）, 比较研究

Abstract:

[Purpose/Significance] Against the global surge of generative artificial intelligence (GenAI) and large language models (LLMs), academic libraries are undergoing a critical paradigm shift in their reference services. While "AI Virtual Librarians" (AIVL) are increasingly adopted to enhance efficiency, cross-national evidence regarding how they are configured alongside traditional "Human Live Reference" (HLR) remains scarce. This study aims to reveal the structural differences in human-AI configurations between Chinese and international top-tier university libraries. It seeks to identify the divergence between "technology-driven" and "human-centric" service models and proposes a governance-oriented hybrid pathway to inform the digital transformation of academic libraries. [Method/Process] The study established two high-resource samples: 42 libraries from China's "Double First-Class" universities and 94 libraries from the U.S. News Top 100 World Universities. A systematic website investigation and standardized interaction tests were conducted to collect data on service availability and deployment models. The study not only quantified the deployment of HLR and AIVL (classified into rule-based and LLM-based) but also qualitatively evaluated the "Core Service Contents" and "Linkage Mechanisms" (e.g., traceability, boundaries, and human fallback). Chi-square tests were employed for statistical analysis, and robustness checks were performed using both broad and strict counting rules to ensure validity. [Results/Conclusions] Results indicate that while the overall service coverage is similar across groups (approx. 74%), the service structure diverges significantly. International libraries predominantly rely on the "Human-only" mode (66.0%), prioritizing deep research support, academic integrity, and privacy protection. In contrast, Chinese libraries show a significantly higher adoption of AIVL (57.1% vs. 8.5%) and LLMs (26.2% vs. 1.1%), with 52.4% operating in an "AI-only" mode. Content analysis reveals that Chinese AIVLs focus on transactional efficiency and 24/7 accessibility, whereas international counterparts focus on distinct research guides and governance. The study identifies a critical trade-off: China's aggressive AI adoption enhances accessibility but faces challenges regarding answer hallucinations and the lack of human fallback mechanisms. To address these challenges, the paper recommends a "Human-AI Collaborative Loop" model. Key strategies include: 1) Implementing risk-tiered routing, where low-risk transactional queries are handled by AI and high-risk research inquiries are directed to humans; 2) Optimizing AI reliability through Retrieval-Augmented Generation (RAG) and controlled knowledge bases to ensure traceability; 3) Establishing clear governance boundaries and stratified implementation paths for libraries with different resource levels, ensuring a balance between technological innovation and service ethics.

Key words: AI virtual librarian, large language models, human-AI collaboration, academic libraries, online reference services, retrieval-augmented generation (RAG), comparative study