The developmental pace of Artificial Intelligence technologies is rapid, with numerous library vendors adding AI tools to their products over the past two years. We will examine a few categories of these features, from tools that translate free text queries into faceted advanced searches to fully-fledged research assistants. With a basic understanding of the types of AI in use at libraries, we will review prominent AI evaluation frameworks. We will investigate library-specific rubrics such as the Ethical AI Assessment Tool from San Diego State University as well as more generic frameworks. Evaluating AI tools is important because they do not come without risk. We will review the major categories of concern: AI data centers consume inordinate amounts of water and electricity; systems trained on biased data tend to reproduce societal biases in their output; AI tools are often employed to undermine or threaten the livelihood of workers; some training data was compiled with massive and systemic copyright violation as evidenced by the Bartz v. Anthropic settlement; the output of AI tools is stochastic, opaque, and occasionally contains "hallucinations"; overeliance on AI can lead to cognitive decay or, in extreme cases, emotional disturbance; and corporate ownership over AI technology has little transparency and regulation. The pertinence of these concerns will vary across libraries and AI platforms. We will filter our evaluatory tactics through the lens of your library's values. In particular, we will heed the amount of labor necessary to evaluate new offerings so we can scale the work appropriately. It is impractical for a small library to navigate a twenty-point rubric every time a vendor announces a new AI-powered feature, but by publishing and reusing evaluations libraries benefit from the strength of our professional community. Finally, this session will also discuss the nascent AI Refusal movement in libraries and consider whether it is right for you. AI Refusal rejects all usage of generative AI technology, citing the numerous harmful consequences mentioned earlier.
Phetteplace et al. (Wed,) studied this question.