Apologizing artificial intelligence: designing and evaluating effective AI apologies after errors

Abstract Artificial intelligence (AI) tools have become increasingly accurate, but still occasionally err. While research started examining AI erring and the efficacy of trust repair attempts in such cases, there is a need to (1) benchmark trust repair attempts by AI against the trust repair dynamics when humans make the same errors, and (2) examine contextual nuances of trust repair dynamics. Using two experiments, we found that users were less forgiving when AI made a noticeable advice error and apologized than when a human expert erred and apologized. Unlike human experts, AI could not restore reliance through a simple apology, and this strategy could even have detrimental effects. The AI’s trust repair was more effective when the AI apologized with an external attribution (i.e., mitigating its accountability) than with an internal attribution (i.e., acknowledging its own limitations). Both effects were stronger in objective tasks than in subjective tasks.