Desiderata for Explainable AI in statistical production systems of the European Central Bank

Explainable AI must be designed for specific user roles, not just theoretically: The paper identifies that different users in statistical production systems (such as data analysts, supervisors, and policymakers at the European Central Bank) have distinct explainability needs, emphasizing that one-size-fits-all AI explanations are insufficient for real-world deployment.

Real-world testing reveals gaps between theory and practice: The authors note that "the usefulness for real-world use cases remains uncertain" despite extensive academic work on explainable AI, demonstrating that solutions evaluated only conceptually may not address actual operational challenges in institutions like central banks.

Explainability is essential for detecting bias and ensuring fairness in high-stakes decisions: The paper establishes explainable AI as "a fundamental step towards establishing fairness and addressing bias in algorithmic decision-making," particularly important in central banking where decisions affect financial systems and public trust.

Practical applications include critical data quality functions: The two concrete use cases—outlier detection in securities databases and identifying data quality checks for banking supervision—show that explainability directly supports essential operational functions that protect financial system integrity.

Specific techniques should be matched to user needs: Rather than recommending universal methods, the paper links different explainability techniques to particular user roles and their specific questions, suggesting that responsible AI implementation requires tailoring explanations to what different stakeholders actually need to understand and act upon.