Preemptively Pruning Clever-Hans Strategies in Deep Neural Networks

User agreement with AI explanations is not a safety guarantee: Even when users and AI systems agree on the reasoning behind a decision, the model may still be relying on hidden, problematic patterns (called "Clever Hans effects") that remain undetected and could cause failures in real-world deployment.

Explainable AI alone has limitations: While explainability tools help identify some model flaws by comparing explanations to domain expert knowledge, they cannot catch all problematic decision-making strategies—particularly those that appear reasonable on the surface but lack genuine understanding.

A new method can proactively reduce hidden biases: The paper introduces EGEM (Explanation-Guided Exposure Minimization), which actively removes model variations that haven't received positive feedback from explanations, rather than waiting for problems to emerge after deployment.

Preemptive pruning improves real-world performance: Models treated with this approach showed both reduced reliance on hidden flawed strategies and higher accuracy on new data, suggesting the method catches problems before they impact users.

Responsible AI requires multiple safeguards: The research demonstrates that responsible AI development needs layered approaches—combining explainability tools with additional verification methods—rather than relying on any single technique to ensure models behave as intended.