Towards Human-centered Explainable AI: A Survey of User Studies for Model Explanations

User studies in XAI are still underdeveloped: Despite explainable AI being considered essential for responsible AI development, the survey found that user evaluations remain "sparse" across the field, meaning most XAI research lacks empirical evidence of whether explanations actually help real people understand AI systems.

XAI adoption is uneven across industries: The research shows that explainable AI is advancing much faster in some domains like recommender systems than in others, suggesting that some high-stakes applications (like healthcare or criminal justice) may lack sufficient human-centered evaluation of their explanations.

Current XAI research ignores human psychology: The survey reveals that user studies in XAI "hardly incorporate any insights from cognitive or social sciences," meaning researchers are often designing explanations without understanding how people actually learn, make decisions, or build trust—limiting the real-world effectiveness of these systems.

Four key factors determine if explanations work: The 97 papers analyzed measured explanations primarily through trust, understanding, usability, and human-AI collaboration performance, providing a framework for practitioners to evaluate whether their explanations are actually serving users' needs.

Practical guidelines are now available: The survey proposes concrete best practices and design guidelines for conducting user studies in XAI, offering researchers and practitioners a roadmap for moving toward more rigorous, human-centered evaluation of AI explanations.