Rules or Weights? Comparing User Understanding of Explainable AI Techniques with the Cognitive XAI-Adaptive Model

Different explanation types work better for different tasks: The research found that when people need to understand why an AI made a decision (forward tasks), they perform differently than when they need to imagine how to change that decision (counterfactual tasks). This suggests organizations should choose explanation methods based on what users actually need to do with the information.

Simpler explanations aren't always better: While linear weights (numerical importance scores) were easier for people to recall and apply than decision tree rules, the study identified 7 different reasoning strategies people use to interpret explanations. This means a one-size-fits-all approach to AI transparency may not work—the best explanation depends on how people naturally think about the problem.

Context matters more than the explanation method itself: The research demonstrated that "the helpfulness of XAI depends on the application data context," meaning the same explanation technique could be clear in one situation but confusing in another. Organizations implementing AI explainability should test their chosen methods with real users and realistic scenarios rather than assuming a technique will work universally.

We now have a framework to evaluate and compare explanation methods: CoXAM provides a cognitive model that can predict how well different explanation techniques will help people understand AI decisions. This gives organizations a more scientific basis for choosing between competing XAI approaches rather than relying on intuition or marketing claims.