CNI Spring 2026 Membership Meeting

Academic libraries are under increasing pressure to deliver fast, high‑quality virtual reference services amid rising demand, staffing constraints, and expectations for always‑available support. Generative artificial intelligence (AI) is frequently marketed as a scalable solution, often without sufficient evidence about where automation might genuinely help or where it may introduce new burdens or risks of suboptimal service. This project briefing details a multi-institutional study of nearly 35,000 anonymized transcripts using a structured natural language processing (NLP) framework. It shares lessons learned from methodological evolution, testing a spectrum of approaches from off-the-shelf interfaces (Microsoft Copilot, Gemini) to a custom-engineered API pipeline. By pairing human expertise with an auditable codebook, the framework generates AI-reasoning trails for every assigned code, allowing researchers to see exactly where model logic aligns with or diverges from professional judgment. The results highlight promising opportunities for responsible automation but also document the extensive iterative cleaning and error-correction cycles necessary to achieve reproducible results. By foregrounding these insights and the often-invisible labor required, the session will provide library leaders with a realistic roadmap for responsible AI integration, moving beyond theoretical promises to offer evidence-based guidance on staffing, workflow design, and long-term strategy in scaling AI-assisted virtual reference.

The University of Texas Libraries (UTL) entered into a charter program with ITHAKA in the fall of 2025 with the objective of testing Seeklight, JSTOR's artificial intelligence (AI)-based metadata generation tool. The goal for the first year of the charter was to test the quality of Seeklight's metadata output and its compatibility with multilingual and complex records. UTL evaluated the quality and changes over time over several months of testing, using born-digital and digitized records sourced from UTL's Benson Latin American Collection and Alexander Architectural Archive. The rubric used during this testing period is a four-point, subjective rating scale, which was developed by UTL's AI Metadata Creation Working Group during earlier tests with out-of-the box LLMs. This presentation will highlight our initial impressions of Seeklight's metadata output on a field-by-field basis. We will explore UTL's continued testing of Seeklight and its current ability to provide the subject knowledge and enhanced discoverability needs for archival records.

As generative artificial intelligence (AI) reshapes how students conduct research, academic libraries face a critical opportunity to redefine their value proposition. This session presents findings from two years of research with over 1,000 undergraduate and graduate students across disciplines, revealing that while AI chatbots have become essential cognitive partners for brainstorming and navigating complex research, students consistently turn to library resources for credibility verification and validation. The data demonstrates that students are thoughtful, skeptical consumers of AI who routinely cross-check AI-generated content against library sources, who view AI outputs, particularly citations and reasoning, with appropriate caution. Rather than displacing libraries, AI has elevated their role as the authoritative validator underlying AI-mediated workflows. This research offers evidence-based insights to inform strategic decisions around discovery systems, research support services, and resource access in an AI-integrated environment, positioning libraries not as competitors to AI tools but as essential stewards of trustworthy, credible knowledge in an era of information abundance.