[Paper Review] On the use of artificial intelligence in financial regulations and the impact on financial stability

Jón Danı́elsson, Andreas Uthemann|arXiv (Cornell University)|Oct 17, 2023

Insurance and Financial Risk Management4 citations

TL;DR

This paper examines how artificial intelligence (AI) can threaten financial stability through misaligned objectives, data limitations, and the uniqueness of systemic crises, proposing six criteria to assess private-sector AI use in financial regulation. It argues that while AI enhances microprudential regulation, its application in macroprudential policy is high-risk due to infrequent, complex crises and opaque decision-making, urging cautious, transparent deployment with human oversight to prevent systemic destabilization.

ABSTRACT

Artificial intelligence (AI) can undermine financial stability because of malicious use, misinformation, misalignment, and the AI analytics market structure. The low frequency and uniqueness of financial crises, coupled with mutable and unclear objectives, frustrate machine learning. Even if the authorities prefer a conservative approach to AI adoption, it will likely become widely used by stealth, taking over increasingly high-level functions driven by significant cost efficiencies and superior performance. We propose six criteria for judging the suitability of AI.

Motivation & Objective

• To analyze how AI adoption in financial regulation may undermine systemic stability despite its efficiency gains.
• To identify key economic and technical challenges—such as data scarcity, crisis uniqueness, and objective misalignment—that limit AI's effectiveness in macroprudential policy.
• To propose six evaluative criteria for assessing the suitability of AI use in financial regulation and crisis resolution by the private sector.
• To examine the risks of AI-driven decision-making in systemic crisis resolution, where data is sparse and outcomes are highly uncertain.
• To highlight the danger of 'stealth' AI adoption by financial institutions, which may outpace regulatory oversight and increase systemic fragility.

Proposed method

• Uses a spectrum framework to assess AI applicability across financial regulation tasks, from high-frequency microprudential to rare, complex macroprudential applications.
• Applies concepts from reinforcement learning and deep learning to model AI decision-making under uncertainty, particularly in crisis scenarios.
• Evaluates generative AI models—such as transformers, diffusion models, and GANs—for simulating financial market scenarios and stress-testing regulatory policies.
• Proposes transfer learning as a method to adapt foundation models to financial contexts using specialized datasets, including central bank rulebooks and economic literature.
• Analyzes the risks of reward hacking and objective misalignment in AI agents through reinforcement learning with human feedback from financial experts.
• Assesses the impact of non-representative, incomplete, or inconsistently measured financial data on AI training and decision quality.

Experimental results

Research questions

• RQ1 How does the infrequency and uniqueness of systemic financial crises challenge the reliability of AI models trained on historical data?
• RQ2 In what ways can misaligned AI objectives or reward hacking undermine financial stability during crisis resolution?
• RQ3 What are the key technical and economic barriers to using AI effectively in macroprudential regulation compared to microprudential regulation?
• RQ4 How does the 'stealth' adoption of AI by private financial institutions threaten regulatory oversight and systemic stability?
• RQ5 To what extent can generative AI models simulate realistic financial market scenarios for stress testing and policy evaluation?

Key findings

• AI poses a significant risk to financial stability when used in macroprudential regulation due to the scarcity of crisis data and the unique, non-repeating nature of systemic events.
• Even with high performance on routine tasks, AI struggles with 'once-in-a-working-lifetime' events like systemic crises, where decision-making opacity and model fragility increase systemic risk.
• Private sector AI adoption is likely to proceed by stealth, driven by cost efficiency and performance, undermining regulatory control even if authorities resist direct AI use.
• Generative AI models such as transformers and diffusion models can simulate financial scenarios useful for stress testing, but their reliability depends on the quality and representativeness of training data.
• Transfer learning and fine-tuning of foundation models on financial rulebooks and expert knowledge can improve AI relevance, but do not eliminate risks from objective misalignment.
• The distinction between AI providing advice and making autonomous decisions blurs when models are opaque, increasing the risk of undetected errors in crisis resolution.