Regulation, compliance and governance of artificial intelligence: the European experience

What it is about

The new European law on the use of artificial intelligence and Europe’s experience in AI regulation, compliance and governance.

Session summary

The two speakers, from a consulting firm with banking experience, explain the European artificial intelligence act, approved in March 2024, which applies to all sectors and defines artificial intelligence very broadly: it includes machine learning, expert systems, decision engines and certain statistical techniques, not only generative artificial intelligence. Its axis is people’s fundamental rights. It classifies systems by risk level (unacceptable, high, limited and minimal): it bans practices such as social scoring or subliminal manipulation, and requires high-risk systems, among them credit, to have transparency, explainability, human oversight, technical robustness, cybersecurity, data quality and logging of all interactions.

They stress a point that matters for banks: if an institution modifies or adapts a third-party tool, it moves from user to provider and takes on a provider’s obligations. On the practical side they discuss how to prevent biases that enter a credit model indirectly, the fact that an advanced capital model built with machine learning has already been approved, and use cases across the credit risk cycle, such as reviewing documents, assisting collections agents and auditing whole populations instead of samples. They close with governance: national authorities and European committees, operating models in production, change management and training of people. The questions covered explainability techniques such as SHAP, provider liability and the lack of gradation when a model is modified.

Topics covered

• What the European act’s definition of artificial intelligence covers
• Risk levels: banned practices, high risk and obligations
• High risk in banking: explainability, human oversight, data and logging
• When an institution moves from user to provider of a tool
• Use cases across the credit risk cycle and bias prevention
• Governance: authorities, operating models and change management
• Questions: explainability techniques, liability and training