Effect of Citations on User Trust in LLM-Generated Answers
An empirical characterization, via a randomized controlled trial, of how citations affect self-reported user trust in LLM chatbot answers. Key results: the presence of citations significantly raises perceived trustworthiness even when the citations are random (though random citations are trusted less than valid ones); trust saturates immediately, with one citation as trustworthy as five; and users who actually check a citation report lower trust, with a checked random citation trusted no more than none at all. The findings operationalize and support the trust-as-anti-monitoring account (checking = reduced trust) alongside the Principle of Social Proof.
Citations and Trust in LLM Generated Answers Yifan Ding1, Matthew Facciani1, Amrit Poudel1, Ellen
A randomized controlled trial (AAAI 2025) measuring how citations shape self-reported user trust in answers produced by a commercial LLM chatbot. Using a between-subjects 3x2 factorial design, 303 Pr…