Towards Feasible Counterfactual Explanations: A Taxonomy Guided Template-based NLG Method

Feasibility matters for AI explanations: The paper demonstrates that many existing counterfactual explanation methods tell users what changes would alter an AI decision without considering whether those changes are actually possible in real life—for example, suggesting someone lower their age to get approved for a loan. This research prioritizes "actionable" explanations that users can realistically implement.

A taxonomy helps communicate complex AI decisions: The researchers developed a four-category system for classifying features based on whether they can be changed (like income), cannot be changed (like age), are partially controllable (like credit score), or are contextual. User studies showed this structured approach helped people better understand what an AI system would need to see differently to make a different decision.

How explanations are worded significantly impacts understanding: Through user studies, the team identified that both the content (which features to mention) and structure (how to phrase the explanation) of counterfactual explanations affect comprehension. Their template-based approach, which carefully considers terminology and presentation, received higher user satisfaction ratings than standard methods across three different application domains.

Ethical considerations should be built into AI explanations from the start: The paper explicitly addresses "immutability and ethical concerns"—recognizing that some features (like race or gender) should never be suggested as changes, even theoretically. This reflects a responsible AI principle that explanation systems must be designed with fairness and ethics in mind, not added afterward.