How should AI decisions be explained? Requirements for Explanations from the Perspective of European Law

Different laws require different types of explanations: The paper finds that European legal frameworks (including the GDPR, AI Act, and product liability laws) each demand distinct explanation properties from AI systems, meaning a one-size-fits-all approach to AI explainability won't satisfy legal requirements across different contexts.

Current AI explanation methods fall short of legal standards: Existing explainable AI techniques don't fully meet legal requirements, particularly regarding "correctness" (fidelity—how accurately explanations reflect what the AI actually does) and confidence estimates (how certain we can be about those explanations), creating a gap between what the law expects and what technology currently delivers.

Explainability requirements span multiple legal domains: Beyond the prominent AI Act, other established European laws—including data protection (GDPR), product safety, and liability frameworks—impose their own explanation requirements on AI systems, suggesting responsible AI governance requires attention to a broader legal landscape than often discussed.

Fiduciary and trust-based standards matter: The paper references "fiduciary plausibility checks" as an international concept, indicating that legal frameworks increasingly expect AI systems to meet standards of trustworthiness and accountability similar to those required of professionals managing others' interests.