Libraries weigh generative AI’s risks and possible uses

At a late-afternoon session hosted during an Institute of Museum and Library Services conference, Arizona State Library project specialist Chris Guerra urged libraries to familiarize themselves with generative artificial intelligence while remaining cautious about its limits and risks. Guerra said the presentation “was written by AI — or was it? Would you know?” to illustrate how indistinguishable machine-generated content can be and why librarians must be prepared to evaluate it.

Guerra outlined the chief concerns libraries should monitor: accuracy and reliability, data privacy and security, bias in training data, misinformation and content appropriateness, intellectual property and copyright, and potential skill erosion among users and staff. He said AI systems “can potentially generate responses that are inaccurate or misleading” and stressed the need for staff to verify outputs.

Rhode Island’s Nicola Buffoni said states are exploring “everyday AI” uses but are approaching implementation conservatively because of public trust and data‑security worries. Buffoni described using chatbots for brainstorming and editing and proposed an idea to build a shared chatbot trained on LSTA subaward project reports so prospective grantees could query prior funded work: “what if we could all build a chatbot trained on LSTA subaward project reports so that prospective grantees can ask questions about grant projects that have already been funded.” She emphasized that outputs must be checked and edited before use.

Speakers recommended practical, role-specific approaches: archivists should focus on preservation and provenance; administrators on budgeting and compliance; technicians on integration and support; outreach staff on communicating ethically with communities; and youth librarians on appropriateness for children. Presenters repeatedly stressed that chatbots and generative tools are aids, not replacements for subject matter expertise.

The session included a conversation about the ethical and legal questions surrounding training data and copyright. Guerra noted that major language models used public datasets such as Common Crawl in their training, which raises questions about rights and attribution. Presenters urged libraries to develop internal policies that protect patron privacy, require human review of AI outputs, and educate patrons about the tools’ limitations.

The session closed with recommendation for hands-on training: practice prompt writing, pilot small use cases (translation, accessibility aids, automated reference), and set clear rules for when to use third-party AI services versus in‑house or vetted tools. The chair then moved participants into table discussions to bring back local impressions and next steps.