Explaining Classifications to Non Experts: An XAI User Study of Post Hoc Explanations for a Classifier When People Lack Expertise

Expertise dramatically affects how people understand AI explanations: The study found that when people were familiar with a domain (MNIST handwritten digits), their understanding of AI explanations changed significantly across multiple measures compared to when they encountered an unfamiliar domain (Kannada digits), suggesting that one-size-fits-all explanations may not work for diverse audiences.

Explanation effectiveness varies by user background, not just explanation type: Response times, perceptions of correctness, and judgments about helpfulness all shifted based on domain familiarity, indicating that the same explanation can be understood very differently depending on whether a user has prior knowledge in that area.

Current XAI research overlooks a critical real-world factor: The authors note that most XAI studies don't account for differences in user expertise, yet this is essential in high-stakes decisions (like medical diagnosis) where trainees and experienced professionals need to understand AI systems differently.

Implications for responsible AI deployment: Organizations implementing AI systems in specialized fields should consider tailoring explanations to users' expertise levels rather than providing generic explanations, particularly in domains where decision-makers have varying levels of experience.