United Kingdom's Prudential Regulation Authority publishes AI and machine learning roundtable slides on SS1/23 model risk management expectations

The Prudential Regulation Authority (PRA) has published the presentation slides from two chief risk officer roundtables held with 21 PRA-regulated firms on 20 and 22 October, focused on firms’ adoption of artificial intelligence and machine learning (AI and ML) in the context of implementing the supervisory expectations in Supervisory Statement 1/23 on model risk management. The slides set out the PRA’s current thinking on key AI and ML model risk themes, including the need for boards to articulate a model risk appetite that addresses the higher uncertainty and opacity of AI and ML models and to link governance arrangements to the model lifecycle through triggers for re-validation. The PRA also flagged observed inconsistencies between firms’ model tiering policies and model inventory submissions, limitations in common explainability and interpretability techniques (including where input features are correlated), and heightened risks around data representativeness and overfitting. Further areas of focus included the suitability of model development testing and independent validation approaches for AI and ML and the adequacy of ongoing performance monitoring, with the slides noting that long monitoring intervals and the absence of quantitative limits may be insufficient for dynamic models and suggesting mitigants such as tracking cumulative changes, pre-approved fallback or challenger models and, in extremis, kill switches. The PRA indicated it will continue engagement through model risk management roundtables and the Artificial Intelligence Consortium, and referenced forthcoming research on explainability and interpretability methods expected to be published in Bank Underground in January 2026.