Allowing humans to interactively guide machines where to look does not always improve human-AI team's classification accuracy

Interactive explanations don't automatically improve decision-making: Despite building an interface (CHM-Corr++) that allowed 18 expert users to interactively edit AI attention maps and see real-time model responses, the study found no statistically significant improvement in classification accuracy compared to static explanations across 1,400 user decisions on bird image classification tasks.

Explainability tools may have limitations beyond their design: The research challenges a common assumption in Explainable AI that simply giving users more interactive control over feature importance maps will enhance human-AI team performance, suggesting that the relationship between understanding and improved accuracy is more complex than previously thought.

Current XAI approaches may need fundamental rethinking: The findings indicate that attention maps and feature importance visualizations—which have become standard in thousands of XAI papers—may not translate into practical benefits for human decision-making, raising questions about whether these explanation methods are optimized for actual human needs.

Expert users still struggle with interactive AI tools: Even domain experts (in this case, bird classification specialists) did not show measurable accuracy improvements when given interactive control, suggesting that interactivity alone is insufficient without addressing deeper questions about how humans process and act on AI explanations.