Reflexive Dialogue-Based Explainability for Human-AI Collaboration: An Empirical Study on Adaptive and Interactive Explanations

Abstract

As artificial intelligence (AI) systems are increasingly integrated into decision-making processes, the need for transparency and user-aligned explanations has become critical. While traditional explainability methods – static or post-hoc – have advanced, they often fail to adapt to users’ evolving informational needs during real-time interactions, limiting their effectiveness in collaborative contexts. This paper addresses this gap by empirically evaluating reflexive dialogue-based explanations, which dynamically adjust explanatory content through iterative user-AI exchanges. We conducted a mixed-method study involving 80 participants performing complex decision-support tasks under two conditions: static explanations and reflexive dialogue. Quantitative results demonstrate that reflexive dialogue significantly improves task accuracy, comprehension, and trust calibration, while qualitative findings reveal enhanced user engagement, perceived agency, and satisfaction. The study identifies key interaction patterns that support cognitive integration and trust refinement. Our main contribution lies in validating a human-centered, adaptive explanation framework that goes beyond one-size-fits-all transparency. The novelty of this work lies in positioning reflexive dialogue not merely as a user support feature but as an essential mechanism for dynamically co-constructing meaning in human-AI collaboration.

Ähnliche Arbeiten