Relevant Irrelevance: Generating Alterfactual Explanations for Image Classifiers

Alterfactual explanations reveal what doesn't matter: Unlike traditional explanations that highlight important features for an AI decision, this research introduces "alterfactual explanations" that show users which aspects of an image can be changed without affecting the classifier's output, helping people understand the full scope of what the AI actually uses to make decisions.

Understanding irrelevance is as important as understanding relevance: The paper argues that knowing which features are irrelevant to an AI's decision is crucial for building accurate mental models of how AI systems work—a gap that traditional explanation methods often leave unfilled.

The approach works on real "black box" AI systems: The researchers successfully demonstrated that alterfactual explanations can be generated for neural network image classifiers without needing access to the model's internal workings, making this technique potentially applicable to AI systems already in use.

User studies confirm complementary value: The paper includes a user study showing that alterfactual explanations work best alongside traditional counterfactual explanations, suggesting that combining both approaches gives people a more complete understanding of AI decision-making than either method alone.

Practical implications for AI transparency: This work addresses a real gap in explainable AI by providing a concrete method to help non-experts understand not just why an AI made a decision, but also what information the AI legitimately ignored—important for building appropriate trust in AI systems.