This paper evaluates whether large-scale language models (LLMs) can serve all users fairly, regardless of demographic characteristics or social status, in a library that leverages virtual reference services. Using six state-of-the-art LLMs, we evaluate whether LLMs discriminate in their responses based on user identity by prompting users to help them in different genders, races/ethnicities, and institutional roles. We find no discrimination based on race or ethnicity, and only a small stereotyped bias toward women in one model. LLMs show subtle adaptations to institutional roles through linguistic choices related to formality, politeness, and domain-specific vocabulary, which reflect professional norms rather than discriminatory treatment.
