User-centric evaluation of explainability of AI with and for humans: a comprehensive empirical study

One-size-fits-all explanations don't work: The study of 39 participants across different expertise levels found that existing XAI methods fail to meet the diverse information needs of different stakeholders—suggesting that AI explanations need to be tailored to whether users are data scientists, domain experts, or general users rather than providing identical explanations to everyone.

Current XAI methods have significant limitations: The research revealed fundamental shortcomings in commonly used explainability algorithms (tested with a Gradient Boosting Classifier), indicating that popular techniques for making AI transparent may not actually help people understand how AI systems make decisions as well as developers assume.

User comprehension should be the gold standard for evaluation: Rather than relying solely on technical metrics, the study demonstrates the importance of empirically testing whether real humans actually understand AI explanations through interviews and comprehension assessments—a social science approach that should become standard practice in responsible AI development.

Domain expertise matters for understanding: Participants with different backgrounds (data science knowledge, visualization skills, domain knowledge about the specific problem) had varying levels of comprehension, highlighting that effective AI explanations must account for what users already know about both AI and the specific domain.

New design principles are urgently needed: The research calls for developing fresh approaches to explainability that go beyond current methods, suggesting the field needs to rethink how explanations are designed from the ground up with actual users in mind.