Assessing Model-Agnostic XAI Methods against EU AI Act Explainability Requirements
New study bridges the gap between technical explainability and Europe's strict new AI regulations.
A new research paper from Francesco Sovrano, Giulia Vilone, and Michael Lognoul tackles a critical challenge for AI developers: aligning technical explainability with legal mandates. The study, titled "Assessing Model-Agnostic XAI Methods against EU AI Act Explainability Requirements," directly addresses the persistent gap between existing XAI techniques and the regulatory expectations set by the landmark EU AI Act. With the Act imposing strict transparency requirements on high-risk AI systems, practitioners have lacked clear guidance on how to achieve compliance, risking market access in Europe.
The researchers' core contribution is a novel qualitative-to-quantitative scoring framework. They analyze model-agnostic XAI methods—tools like SHAP, LIME, or counterfactual explanations that can be applied to any AI model—and relate their interpretability features to the AI Act's specific provisions. Expert assessments of an XAI method's properties are systematically aggregated to produce a regulation-specific compliance score. This framework provides a practical tool for teams to evaluate whether their chosen explanation techniques meet regulatory muster.
Beyond offering a compliance checklist, the paper serves as a crucial map of the current landscape. It helps practitioners identify which XAI solutions may adequately support legal explanation requirements while simultaneously flagging technical shortcomings and ambiguities in the regulation itself. The work, accepted at the 2026 World Conference on eXplainable Artificial Intelligence, underscores that achieving true compliance will require iterative dialogue between technologists, legal experts, and regulators to close the remaining gaps.
- Proposes a scoring framework to translate expert assessments of XAI methods into an EU AI Act compliance score.
- Focuses on model-agnostic XAI techniques (e.g., SHAP, LIME) applicable to any AI model for broad relevance.
- Highlights specific technical issues in XAI that require further research and regulatory clarification for full compliance.
Why It Matters
Provides a crucial roadmap for AI companies to navigate EU compliance, turning legal requirements into actionable technical assessments.