Examining How Explanations from Multiple AI Agents Affect User Trust After AI Refusal

Abstract

AI systems are designed to decline certain requests to avoid providing harmful or inaccurate information. However, refusals can unintentionally reduce users’ impressions of and trust in AI. Prior work has examined what an AI should explain when refusing a request, but offers no guidance on which AI should deliver the explanation when multiple agents are involved. Using a multi-agent scenario, we compared two explanation conditions, delegated explanation by a third-party AI on behalf of the primary AI and shared explanation by both AIs, against two complementary baselines: no explanation and self-explanation by the primary AI. Results showed that users’ impressions of the primary AI dropped when it delegated the explanation about the refusal to the third-party AI. Moreover, this negative impression transferred to the third-party AI: even without directly refusing, it was perceived as reinforcing the primary AI’s refusal. We discuss challenges and opportunities in designing multi-agent AI explanations for refusing users’ requests.