Even-Ifs From If-Onlys: Are the Best Semi-Factual Explanations Found Using Counterfactuals As Guides?

Semi-factual explanations ("even-if" statements) may be more useful than counterfactuals for helping people challenge unfavorable AI decisions. While counterfactuals explain what changes would flip a negative decision, semi-factuals show what features can stay the same while preserving that decision—potentially offering more realistic and achievable paths forward for affected individuals.

Counterfactual guidance is not necessary for generating high-quality semi-factual explanations. The study tested 8 different semi-factual methods across 7 datasets and found that methods that don't rely on counterfactuals as a starting point actually performed comparably or better than those that do, suggesting simpler approaches may be equally effective.

Different aspects of how an AI system makes decisions matter more than using counterfactuals as a guide. Rather than building semi-factual methods on top of counterfactual analysis, the research indicates that directly examining other features of the decision-making space produces better explanations for users.

This work highlights the importance of empirically testing XAI methods before deployment. By comprehensively comparing 8 methods across multiple datasets and metrics, the authors demonstrate that intuitive assumptions about how to build explanation systems (like using counterfactuals as guides) don't always hold up in practice—a critical lesson for responsible AI development.