Explainability (explainable AI)

Explainability refers to the ability to make an artificial intelligence model's decisions or predictions understandable to a human being. It stands in contrast to the black-box character of the most complex models, whose performance is often accompanied by an opacity such that even their designers struggle to explain why a precise decision was taken. This tension between performance and transparency is one of the central dilemmas of applied AI, the most powerful models frequently being the least interpretable. Explainability is not merely a technical concern, it has become a major regulatory and legal issue. Data protection law frames fully automated decisions and recognizes a right to information about the underlying logic, and the new frameworks applicable to AI impose transparency requirements for high-risk systems. Dedicated techniques attempt to shed light, after the fact, on how models work. For insurance and liability, explainability is decisive, because it is hard to attribute fault, to contest a decision or to compensate harm when no one can explain what happened inside the model, opacity diluting responsibilities and complicating proof.

A bank refuses a loan on the recommendation of a model whose decision it cannot explain. The customer contests it, the regulator queries it, and the lack of explainability places the institution in difficulty in justifying a refusal that nonetheless carries heavy consequences.

explicabilité, IA explicable, explainable AI, XAI, interprétabilité