Generative AI Model Risk Management Framework

This paper is a technical report that presents a framework for model risk management (MRM) specifically for Large Language Models (LLMs) in the finance and insurance sectors. It discusses the need to reconsider existing MRM principles, metrics, and prioritization due to the dynamic nature of LLMs, which generate novel content and require continuous monitoring rather than static pre-deployment validation. The report advocates for a shift towards human-in-the-loop (HITL) monitoring of AI systems, while also highlighting the psychological challenges faced by humans in low-event rate monitoring situations. It proposes leveraging AI-assisted compliance monitoring to ensure accuracy and adherence to regulatory and ethical standards. The paper emphasizes the importance of disaster planning, establishing clear metrics for AI disconnection, and the necessity of robust fallback plans. It concludes that while traditional MRM principles apply, their application must be rethought to effectively integrate Generative AI into critical operations while managing associated risks.