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

Ioannis Papantonis^*Vaishak Belle

• School of Informatics, University of Edinburgh, Edinburgh, United Kingdom

As AI and ML models integrate into high-stakes fields like healthcare and criminal justice, fully automated decision-making can introduce ethical risks, highlighting the need for a balanced human-AI collaborative approach. However, achieving effective human-AI collaboration requires that users not only trust the model's outputs but also understand when and why they should rely on them. Through an empirical evaluation, this work assesses whether combining uncertainty estimates and explanations enhances the effectiveness of interactions beyond the benefits offered by either component alone, particularly in relation to users' self-confidence. Findings indicate that while uncertainty estimates alone can enhance accuracy, explanations significantly increase model understanding, demonstrating complementary strengths. Furthermore, the study shows that users' self-confidence influences reliance, understanding, and trust, underscoring the complex dynamics in human-AI interactions. Together, these insights provide a pathway for designing AI systems that foster trust, enhance understanding, and promote effective, ethically grounded collaboration in critical applications.