I-CEE: Tailoring Explanations of Image Classification Models to User Expertise

One-size-fits-all explanations may not work for everyone: The paper demonstrates that current AI explanation techniques typically provide identical explanations to all users, regardless of their background or expertise level. I-CEE addresses this by tailoring explanations based on whether users are technical experts or non-experts, suggesting that personalized explanations could improve how well people understand AI decisions.

Tailored explanations improve users' ability to predict model behavior: In experiments with simulated users, I-CEE improved "simulatability"—the ability to accurately predict what the model would decide in new situations—compared to standard explanation approaches. This suggests that matching explanation complexity to user expertise helps people develop better mental models of how AI systems work.

Effective AI explanations require understanding the user, not just the model: The research emphasizes a "human-centered" approach to explainable AI that considers who is receiving the explanation. Rather than focusing solely on technical explanation methods, the paper argues that responsible AI deployment requires considering the explainee's background and needs.

Example-based explanations with local context are valuable: I-CEE explains model decisions by showing users relevant training examples alongside specific explanations and the model's reasoning. This multi-faceted approach combining examples, local explanations, and model decisions may be more effective than single explanation types alone.

Human evaluation is essential for validating explanation methods: The research included experiments with 100 human participants, not just simulated users, highlighting that real-world testing with actual people is necessary to determine whether explanation techniques truly improve understanding and trust in AI systems.