Can Conversational XAI Improve User Performance? An Experimental Study

Conversational AI explanations don't automatically improve performance: Despite promises that conversational XAI assistants would help users work better with AI models, the study found no significant performance difference between conversational assistance and simpler Q&A-based assistance (N=42), suggesting that the format of explanation alone may not be the key factor in helping users.

Users can outperform AI models when given the right tools: Participants in both assistance conditions significantly outperformed the underlying prediction model by identifying and compensating for its systematic errors, indicating that providing explanations—regardless of format—can empower users to catch and correct AI mistakes.

Current engagement with XAI tools remains modest: The study noted "modest engagement overall" with the explanation assistance, raising concerns about whether users are actually utilizing the explanations available to them, even when those explanations could help them perform better.

Explanation design alone is insufficient: The findings suggest that simply making AI systems "explainable" or adding conversational interfaces isn't enough to improve real-world user performance; future work needs to focus on engagement strategies and understanding why explanations help when they do.

More research needed on explanation mechanisms: The authors plan to investigate "the mechanisms driving improved predictions," acknowledging that we still don't fully understand how and why explanations help users work effectively with AI systems.