What does EBA follow-up report on machine learning for IRB models (EBA/REP/2023/28) say about AI in banking?

What are the key points of EBA follow-up report on machine learning for IRB models (EBA/REP/2023/28)?

• Published August 4, 2023 as EBA/REP/2023/28.
• Current use is selective: ML in core IRB modelling steps, ML as challenger models for validation, and collateral valuation, rather than fully ML-based rating systems.
• Complexity issues addressed: overfitting, human-skill requirements, and explainability of individual outputs.
• Reiterates the discussion paper's recommendations on understanding, documentation, validation and stability, which respondents broadly supported.
• Discusses interaction with GDPR (automated decision-making) and the AI Act, calling for clarification to avoid unintended consequences for IRB models.
• No new guidelines were issued; the report is the standing statement of EBA expectations for ML in capital models.

What did EBA follow-up report on machine learning for IRB models (EBA/REP/2023/28) change for banks?

The follow-up turned a discussion into a settled position: ML in IRB models is acceptable within existing CRR requirements, and banks should expect supervisors to test explainability, understanding and stability rather than prohibit the technique. It also put on record, a year before the AI Act was adopted, that credit-scoring models used for capital purposes could be caught by the Act — an issue the EBA's 2025 mapping exercise later addressed.