Leveraging Mental Models for Human-Centered Explanations: A Systematic Design Approach Illustrated by Explainable Fake Review Identification

Mental models matter more than technical accuracy: The paper found that explanations aligned with how people naturally think about problems (their "mental models") were significantly more understandable and useful than explanations that simply showed which parts of the AI's input it focused on, even when the latter were technically more transparent.

Users want commonsense reasoning, not technical details: In studying how people assess fake reviews, researchers discovered that people rely on everyday judgment heuristics (like checking if a review seems rushed or uses suspicious language patterns) rather than statistical measures—and explanations based on these commonsense concepts were rated as more satisfying and trustworthy by 166 study participants.

Current XAI methods often fail users despite appearing transparent: The paper demonstrates that popular "token-level transparency" methods used in AI text classification—which highlight individual words the model focused on—don't actually help people understand or trust the AI's decisions as well as explanations grounded in human reasoning patterns.

Better explanations improve both understanding and accountability: When explanations were designed using a systematic process (MIND-X) that incorporated user research and mental models, participants reported better decision-making support and perceived the AI system as more trustworthy and persuasive, suggesting this approach could support more responsible AI deployment.

One-size-fits-all explanations are insufficient: The research emphasizes that effective XAI requires understanding the specific context and user group, rather than applying generic transparency techniques—a finding with implications for how organizations should design AI explanations for different audiences.