Selective Explanations: Leveraging Human Input to Align Explainable AI

Human explanations are selective by nature: The paper identifies a key gap between how AI systems currently explain themselves and how humans naturally communicate—humans typically highlight only the most relevant reasons for a decision rather than listing everything, yet most XAI tools present exhaustive lists of factors.

Personalized explanations improve decision-making: The research found that when AI explanations were tailored to what individual decision-makers considered relevant (either through their own input or preferences from similar users), people made better decisions and were less likely to over-rely on the AI system's recommendations.

Two practical approaches to get human input: The study tested both direct methods (asking users to specify what matters to them) and indirect methods (using preferences from panels of similar users), suggesting organizations have flexible options for implementing selective explanations without requiring extensive individual customization.

Selective explanations enhance user trust and perception: Beyond improving actual decision quality, participants reported more positive views of the AI system when explanations were tailored to their priorities, suggesting that alignment between how AI explains itself and what humans care about affects both trust and usability.

Small amounts of human feedback can go a long way: The framework requires only "a small sample" of human input to generate better explanations for many users, indicating that organizations don't need massive annotation efforts to make XAI systems more effective and aligned with human needs.