The Utility of "Even if..." Semifactual Explanation to Optimise Positive Outcomes

Semifactuals offer a new perspective on AI explanations: Rather than only explaining how to fix negative decisions (e.g., "earn more to get approved"), this research introduces "semifactual" explanations that show flexibility within positive outcomes (e.g., "you'd still be approved even if you reduced your down payment"). This represents a shift from problem-solving to opportunity optimization.

User studies confirm practical value: When people received positive outcomes from an automated system, they found semifactual explanations more useful than traditional counterfactual explanations, suggesting this approach better meets real user needs and could improve trust in AI systems that make favorable decisions.

Causality matters for meaningful explanations: The research demonstrates that using causal reasoning—rather than just statistical correlations—is important for generating semifactual explanations that accurately reflect how systems actually work, which has implications for the reliability of AI explanations more broadly.

"Gain" quantifies user benefit: The paper introduces a measurable concept of "gain" to evaluate how much users benefit from explanations, providing a concrete way to assess whether explanations actually help people make better decisions or understand their options more clearly.

Responsible AI should address positive outcomes too: By focusing on optimizing positive decisions rather than only correcting negative ones, this work highlights an underexplored area in responsible AI—ensuring that when systems make favorable decisions, users understand the flexibility and robustness of those decisions.