Counterfactual Explanation
An explanation that describes the minimum changes to input features required to change a model's decision — answering 'what would have to be different for a different outcome?'.
Counterfactual explanations are the most actionable form of AI model explanation — they tell affected individuals not just why a decision was made, but what they would need to change to get a different outcome.
Example: A loan applicant is rejected. A SHAP explanation says "your debt-to-income ratio had the highest negative impact." A counterfactual explanation says "if your monthly debt payments were reduced by ₹8,000, you would likely be approved."
Wachter et al. (2017) formalised counterfactual explanations in the context of GDPR's "right to explanation" — arguing that counterfactuals are the legally appropriate form of explanation for automated decisions, as they are actionable and do not expose model internals.
Counterfactual Fairness (Kusner et al., 2017): A stricter fairness criterion requiring that a model's decision for an individual would be the same in a counterfactual world where their protected attributes (race, gender) were different. This is a test for whether protected attributes causally affect outcomes — as distinct from merely being correlated with predictive features.
Generation methods: Finding the closest counterfactual input is an optimisation problem. Common approaches include: gradient-based search (for differentiable models), genetic algorithms, DICE (Diverse Counterfactual Explanations), and constraint satisfaction.
Limitations: Not all counterfactuals are actionable (a rejected applicant cannot change their age). Good counterfactual systems only propose changes to mutable, actionable features.