Counterfactual Explanation

An explanation technique that communicates what minimal change to the input would have produced a different output from an AI model, for example 'if the applicant's income had been $5,000 higher, the loan would have been approved'. Counterfactual explanations are legally relevant under regimes like the GDPR's right to an explanation and are often more intuitive than feature-importance methods. For accessibility, counterfactuals can be rendered purely in natural language without visualisation, making them a promising candidate for accessible XAI, but they can still require cross-checking against the input modality (image, audio) that the disabled user may not be able to perceive.