Previous research presents conflicting findings on how disclosing AI identity affects user responses. This meta-analysis quantifies the overall impact of AI identity disclosure on user responses (evaluations, attitudes, intentions and behaviors). It further explores how these effects are moderated by nine variables across four dimensions.
A systematic literature search identified 33 relevant articles comprising 44 independent studies. From these, 67 effect sizes from 25,208 participants were synthesized using a three-level random-effects model to account for nested data structures. We further employed a series of univariate three-level mixed-effects meta-regression models to examine subgroup differences, followed by exploratory analyses to test interaction effects among moderators, complemented by publication bias assessments and a series of robustness and sensitivity analyses.
AI identity disclosure has a small, statistically significant negative overall effect on user responses. High heterogeneity was observed, with no significant publication bias indicated. Negative effects were stronger for objective responses (vs. subjective responses), non-HCI tasks (vs. HCI tasks) and knowledge-oriented (vs. experience-oriented) applications. Furthermore, exploratory interaction analyses suggested significant interaction effects between sample types and experimental methods, cultural backgrounds and task forms and cultural backgrounds and application scenarios.
This meta-analysis systematically quantifies the impact of AI identity disclosure across diverse user responses, addressing prior inconsistencies by identifying significant contextual moderators. The findings inform context sensitive AI identity disclosure strategies, advance AI ethics and guide human-AI interaction design.
