Contrastive Explanations That Anticipate Human Misconceptions Can Improve Human Decision-Making Skills

TL;DR

This work addresses deskilling risks in AI-assisted decision-making by proposing a human-centered contrastive explanation framework that contrasts AI recommendations with predicted human reasoning. The authors implement a four-module system (AI task model, human foil model, contrast module, and presentation module) to generate contrastive explanations and test them in a large online study (N=628) across five conditions. Results show that contrastive explanations, especially with predicted foils, significantly improve human learning without reducing decision accuracy, while timing and foil quality modulate subjective experiences and learning gains. The findings suggest that AI tools designed to align with users' mental models can upskill decision-makers, offering practical implications for designing explainable AI that supports long-term competence and autonomy in AI-assisted tasks.

Abstract

People's decision-making abilities often fail to improve or may even erode when they rely on AI for decision-support, even when the AI provides informative explanations. We argue this is partly because people intuitively seek contrastive explanations, which clarify the difference between the AI's decision and their own reasoning, while most AI systems offer "unilateral" explanations that justify the AI's decision but do not account for users' thinking. To align human-AI knowledge on decision tasks, we introduce a framework for generating human-centered contrastive explanations that explain the difference between AI's choice and a predicted, likely human choice about the same task. Results from a large-scale experiment (N = 628) demonstrate that contrastive explanations significantly enhance users' independent decision-making skills compared to unilateral explanations, without sacrificing decision accuracy. Amid rising deskilling concerns, our research demonstrates that incorporating human reasoning into AI design can foster human skill development.

Figures (12)

• Figure 1: A simplified illustration of (a) unilateral explanations, which list all the features contributing to the AI's decision, and (b) contrastive explanations, which highlight the differences between the AI's choice and a likely human response for an exercise recommendation task.
• Figure 2: The Contrastive Explanation Framework: The AI task model predicts the AI’s response for a given decision task (fact), while the human model predicts the user’s response for the same task (foil). The contrastive module then analyzes the differences between the AI's and the human's responses, generating task dimensions where the fact is superior to the foil ($S_\textrm{fact}$) and, if any, where the foil is superior to the fact ($S_\textrm{foil}$). Finally, the presentation module, powered by a large language model (LLM), formats the information into an interpretable explanation, filling in small common-sense knowledge gaps within the constraints of the predictive models. The example generation outlined in the figure is relates to the character vignette in Figure \ref{['fig:exercise-task']}.
• Figure 3: Illustration of the exercise recommendation decision-making task featuring different explanation designs. \ref{['fig:contrastive']} shows a sample of the task with contrastive explanation, whereas \ref{['fig:single-sided']} and \ref{['fig:contrastive-after']} depict only the explanations for the respective conditions. In the contrastive random condition, the presentation was identical to the contrastive condition, but with the alternative (foil) selected randomly. In the no-AI condition (not illustrated), participants made decisions without any AI assistance.
• Figure 4: Main results. Marginal means of human learning (post-intervention performance, controlled for pre-intervention performance) and accuracy accross different conditions. Error bars represent one standard error. Significance levels after Holm-Bonferroni correction are presented only for significant (or marginally significant) differences, indicated by: * p < 0.05, ** p < 0.01, *** p < 0.001.
• Figure 5: Overreliance across conditions. Error bars represent one standard error.