Explainability as statistical inference