Assessing Model-Agnostic XAI Methods against EU AI Act Explainability Requirements

Current XAI methods fall short of legal requirements: The paper identifies a "persistent gap" between existing explainability techniques and the EU AI Act's regulatory demands, meaning companies cannot simply apply standard XAI tools and assume they meet legal obligations for transparency.

Model-agnostic methods need evaluation against specific regulations: The researchers assessed XAI approaches (those working across different AI systems) and found that their usefulness depends on how well they address particular legal requirements—there is no one-size-fits-all solution for compliance.

A new scoring framework can guide compliance decisions: The paper proposes a practical tool that converts expert assessments of XAI features into compliance scores, helping practitioners determine which explanation methods might actually satisfy EU AI Act requirements rather than guessing.

Technical limitations require further development: The study highlights that some aspects of the AI Act's explainability requirements cannot yet be fully met by current technology, signaling that both XAI research and regulatory guidance need to evolve together.

Practitioners need clearer guidance on implementation: By mapping XAI methods to specific legal requirements, the paper addresses a real gap where companies operating in the EU market lack concrete direction on how to make their AI systems compliant with transparency rules.