AI explanations that change depending on which model you ask are not much use to anyone.
A paper posted to arXiv (arXiv:2501.05795) tackles a known weakness in counterfactual explanations — the "what would have changed the outcome" answers that AI systems offer to justify decisions. The problem: when several models achieve similar accuracy, counterfactual explanations can vary wildly between them, making them unreliable as a basis for action. The authors introduce a method that borrows the concept of Pareto improvement from economics, using multi-objective optimization to generate explanations that hold up across that spread of competing models. Tests on both simulated and real datasets showed the approach was robust and practical.
Counterfactual explanations are one of the more intuitive tools in the explainability toolkit — tell someone what they would need to change to get a different answer, and they have something actionable. The catch exposed here is that "actionable" advice built on a fragile explanation is worse than no advice at all, especially in high-stakes settings where a decision-support system might be retrained or swapped out without warning. Grounding robustness in social welfare concepts rather than single-model accuracy is an unusual framing that could push the field toward treating explanation stability as a first-class requirement.
Most explainability research still optimizes for a single model in isolation; this work is a reminder that production environments rarely look that clean.