Counterfactual Explanations for AI Decisions

A counterfactual explanation answers a contrastive question: what small change to the input would make this model produce a different output? Wachter, Mittelstadt, and Russell proposed counterfactual explanations as one way to communicate automated decisions without disclosing the system’s full internal logic. For example, an explanation might state that a loan score would cross a threshold if a reported balance were lower, or that an application would score differently if a missing credential were included. The explanation is about the model’s behavior, not necessarily the real-world cause of an outcome. It may identify a feature that is correlated with the model output rather than a factor a person can change to improve their circumstances. “If income were higher” does not mean raising income is easy or that the lender would approve the application after other checks. A counterfactual can also be implausible, unlawful, unfair, or based on inaccurate input data. If it suggests changing protected or immutable characteristics, it may expose a serious design problem. Useful counterfactuals should be valid for the model, close enough to be understandable, and feasible and actionable for the person where the goal is recourse. They should respect constraints among features: increasing age by a decade while holding employment history fixed may describe an impossible case. Multiple counterfactuals can show alternative paths, but each should be checked for realism, stability, and policy constraints. The result should identify the model version and decision context. Counterfactuals do not reveal every reason inside a black box and should not be framed as a legal explanation unless applicable law requires it and the output meets that standard. Use them as one communication and debugging tool. Pair them with data access, error correction, human review, and appeal routes. Test whether users understand the explanation and whether proposed changes are within their control before relying on it for consequential decisions.