Even if Explanations: Prior Work, Desiderata & Benchmarks for Semi-Factual XAI

Semi-factual explanations offer a different perspective than standard counterfactuals: While counterfactuals tell people what would need to change for a different outcome (e.g., "get a shorter loan term"), semi-factuals explain why the current outcome occurred despite certain factors (e.g., "even if you had a shorter loan term, you still would have been denied"). This distinction matters because it helps users understand the robustness of AI decisions.

Semi-factuals have been understudied in AI despite cognitive science research: The paper notes that cognitive scientists have extensively studied how humans naturally use semi-factual reasoning, but AI researchers have largely overlooked this explanation type. Bridging this gap could make AI explanations more aligned with how people actually think about decisions.

There's a need for clear standards in semi-factual explanations: The paper establishes "key desiderata" (desired qualities) for semi-factual XAI systems, suggesting the field currently lacks consensus on what makes a good semi-factual explanation. This standardization is important for ensuring explanations are actually useful and trustworthy to end users.

Existing algorithms need evaluation against benchmarks: The paper tests historical algorithms and introduces new methods, providing benchmark comparisons that can guide future development. This empirical grounding helps move semi-factual XAI from theoretical discussion toward practical, measurable improvements in how AI systems explain themselves.