AUTHOR=Papantonis Ioannis , Belle Vaishak 

TITLE=Why not both? Complementing explanations with uncertainty, and self-confidence in human-AI collaboration

JOURNAL=Frontiers in Computer Science

VOLUME=Volume 7 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/computer-science/articles/10.3389/fcomp.2025.1560448

DOI=10.3389/fcomp.2025.1560448

ISSN=2624-9898

ABSTRACT=IntroductionAs AI systems integrate into high-stakes domains, effective human-AI collaboration requires users to be able to assess when and why to trust model predictions. This study investigates whether combining uncertainty estimates with explanations enhances human-AI interaction effectiveness, particularly examining the interplay between model uncertainty and users' self-confidence in shaping reliance, understanding, and trust.MethodsWe conducted an empirical study with 120 participants across four experimental conditions, each providing increasing levels of model assistance: (1) prediction only; (2) prediction with corresponding probability; (3) prediction with both probability and class-level recall rates; and (4) all prior information supplemented with feature importance explanations. Participants completed an income prediction task comprising of instances with varying degrees of both human and model confidence levels. In addition to measuring prediction accuracy, we collected subjective ratings of participants' perceived reliance, understanding, and trust in the model. Finally, participants completed a questionnaire evaluating their objective model understanding.ResultsUncertainty estimates were sufficient to enhance accuracy, with participants showing significant improvement when they were uncertain but the model exhibited high confidence. Explanations provided complementary benefits, significantly increasing both subjective understanding and participants' performance with respect to feature importance identification, counterfactual reasoning, and model simulation. Both human confidence model confidence played a role in shaping user's reliance, understanding, and trust toward the AI system. Finally, the interaction between human and model confidence determined when AI assistance was most beneficial, with accuracy gains occurring primarily when human confidence was low but model confidence was high, across three of four experimental conditions.DiscussionThese findings demonstrate that uncertainty estimates and explanations serve complementary roles in human-AI collaboration, with uncertainty estimates enhancing predictive accuracy, and explanations significantly improving model understanding without compromising performance. Human confidence acts as a moderating factor influencing all aspects of human-AI interaction, suggesting that future AI systems should account for user confidence levels. The results provide a foundation for designing AI systems that promote effective collaboration in critical applications by combining uncertainty communication with explanatory information.