Alterfactual Explanations -- The Relevance of Irrelevance for Explaining AI Systems

Understanding what doesn't matter is as important as understanding what does: The paper demonstrates through user studies that people develop more accurate mental models of AI systems when they learn not just which features influence a decision, but also which features the AI ignores. This challenges the assumption that explanation methods should focus exclusively on important factors.

Current explanation methods may give incomplete pictures: Traditional counterfactual explanations (showing "what if this feature changed?") only highlight relevant information. The research suggests these approaches leave gaps in user understanding by not explicitly showing users which input characteristics are irrelevant to the AI's decision-making.

Alterfactual explanations show AI robustness to irrelevant changes: By presenting alternative scenarios where irrelevant features are altered while keeping the decision the same, this new approach helps users see which aspects of input data could change arbitrarily without affecting outcomes—providing insight into how stable or brittle an AI system's reasoning actually is.

User studies confirm the method's effectiveness: The paper includes extensive user testing showing that alterfactual explanations "significantly contribute to participants' understanding of an AI" and convey understanding of different reasoning aspects than established methods, suggesting this is a practical improvement over current explanation techniques.

This approach supports more responsible AI deployment: By helping non-experts understand both what an AI considers and what it ignores, alterfactual explanations could enable better human oversight and help identify potential biases or unexpected system behaviors before deployment.