Learning About LLM Biases Impacts Analytical and Emotional Trust of AI Differently for WEIRD vs. non-WEIRD Populations
Résumé
Methods for improving human trust in Large Language Models (LLMs) have been extensively studied in recent research, but limited in their focus on improving users analytical trust in AI, with participants collected primarily from Western, Educated, Industrialized, Rich, and Democratic (WEIRD) backgrounds. These limitations have important implications for AI Ethics, how AI design differently impacts specific groups, and how this can be better accounted for in the future. As LLMs become embedded in consequential decisions in healthcare, education, and employment, the question of how trust forms across different populations and across different types of trust becomes ever more relevant. We report results from an online experiment in which 504 participants completed a pre and post trust questionnaire that distinguished between analytical trust and emotional trust separated by an educational platform on LLM biases. The participant pool was balanced across age and gender with 31.2% classified as WEIRD and 68.8% as non-WEIRD. For WEIRD participants only, education on LLM biases reduced analytical trust significantly more than emotional trust (d=0.301, p=.001), with a significant group effect on this analytical-emotional trust change contrast (t=−2.077, p=.039). Mediation analyses confirmed that no single constituent demographic of the WEIRD classification accounted for this pattern, and a sentiment and individualism–collectivism analysis of participant open responses showed that negative and collectivist responses were associated with trust change profiles more similar to non-WEIRD participants. The findings argue for cross-cultural evaluation of trust in AI and for treating emotional trust as a distinct construct.
Mots-clés
Aucun mot-clé renseigné.