Roadmap of Designing Cognitive Metrics for Explainable Artificial Intelligence (XAI)

Understanding matters more than just having explanations: The paper emphasizes that XAI systems should be evaluated based on whether stakeholders actually understand the explanations provided, not simply whether explanations exist. This shifts focus from technical sophistication to real-world comprehension.

Seven key dimensions should guide XAI evaluation: The research identifies seven critical areas for assessing AI explanations—goodness, satisfaction, user understanding, curiosity & engagement, trust & reliance, controllability & interactivity, and learning curve & productivity—suggesting that responsible AI requires measuring multiple aspects of human experience, not just accuracy.

Psychological and cognitive science methods are essential: The paper recommends incorporating subjective measures (like user surveys) and objective psychological measures to evaluate XAI systems, indicating that responsible AI development requires collaboration between AI researchers and cognitive scientists to properly assess whether explanations work for real people.

Trust and reliance need careful measurement: By identifying "trust & reliance" as a distinct dimension, the research highlights that stakeholders might trust an AI system's explanations without truly understanding them—a critical distinction for ensuring responsible deployment in high-stakes domains like healthcare or criminal justice.

User engagement and learning should be tracked: The inclusion of "curiosity & engagement" and "learning curve & productivity" suggests that good AI explanations should actively help users learn and remain engaged, rather than passively presenting information, making the human-AI interaction more interactive and educational.