When AI models malfunction, address the problem not the math

Governance of artificial intelligence models should focus on actionable outcomes rather than interpretability, argues former chief regulator

When a bank’s value-at-risk model fails backtesting, risk managers know what to do. They trace the problem to specific components – like a stale risk factor time series, or an underperforming revaluation model – and make targeted fixes. This decomposability has been fundamental to traditional model risk management for decades.

But large language models (LLMs) and generative AI systems don’t work