BoE speaks on AI and financial stability

Sarah Breedon spoke at a conference on the opportunities and challenges of emerging technologies in the financial ecosytem. She delved into the novel features of Generative AI and what it might mean for financial stability. Key points included:

• that generative AI models can learn and evolve autonomously and at speed, with outputs that are not always interpretable or explainable, and objectives that might not be clear and may not be fully aligned with society’s ultimate goals;
• at a microprudential level, regulators need to be confident that technology-agnostic frameworks are good enough to mitigate the financial stability risks AI poses – and that the managers of regulated firms can understand and manage what their AI models are doing;
• at a macroprudential level, how to manage the risks of the financial system becoming more dependent on shared AI technology and infrastructure systems;
• at the moment, the BoE does not think it needs to changes its approach;
• the BoE is launching an AI Consortium of the private sector and AI experts and the FPC will publish an assessment of AI’s impact on financial stability.

The latest of 5 annual surveys the UK regulators have been sending to firms comprised questionnaires to nearly 120 firms from various parts of the regulated community. The results will be published soon, but headlines include that now 75% of firms are using some form of AI in their operations – an increase of 22% from 2022. 17% of use cases are using foundation models. 41% of respondents said they were using AI to optimise internal processes and 26% were using it to enhance customer support. Surprisingly, significantly less than half the firms surveyed are using AI as a financial crime prevention tool.

She discussed the challenges that AI models present:

• that models can be dynamic so that the way they behave over time can become misaligned with the original intention;
• a potential lack of explainability;
• the breadth of data on which they are trained;
• users coalescing around a small number of common models; and
• that models are autonomous and so could determine outcomes and make decisions without a senior manager being involved.