Explanations That Work: A Risk-Based Model for Meaningful AI Transparency in Regulated Decisions

Abstract

Law increasingly gives people a right to an explanation when artificial intelligence (AI) contributes to decisions about them. The EU AI Act grants a right to a clear and meaningful explanation for decisions informed by high-risk systems, the Court of Justice of the European Union has held that handing over an algorithm or formula does not satisfy the right to meaningful information about automated decisions, and the UK has reformed its rules on automated decision-making while keeping rights to information and to contest. Yet the empirical literature on explainable AI shows that explanations frequently fail the people they are meant to serve. They can increase acceptance of AI advice whether or not it is correct, leave people less able to spot model errors, and be produced to satisfy compliance rather than to inform. This paper asks what makes an explanation work, and how the effort put into explanation should scale with the risk of the decision. It synthesises legal, regulatory and human-computer interaction research to propose the Meaningful Explanation Model (MEM). The model specifies explanations along three dimensions: the risk tier of the decision, the recipient and the purpose the explanation must serve, and a set of effectiveness tests that judge an explanation by what it enables rather than by whether it exists. The model is applied to five regulated decisions and its implications for regulators, deployers and designers are set out.