Trustworthy AI in Central Banking: International Practices and Financial Stability Risks

Abstract

Yue Dai

Artificial intelligence (AI), and especially recent advances in large language models and foundation models, is becoming relevant to central banking in two distinct ways. It is a tool that can improve macroeconomic analysis, supervisory technology, payment oversight, anti-moneylaundering monitoring, cyber resilience, and financial stability surveillance. It is also a new source of risk that central banks must understand as stewards of monetary and financial stability. This paper reviews emerging international practices in the use of AI by central banks and financial authorities, and connects those practices to the computer science literature on trustworthy AI and AI safety. The central argument is that, for central banks, trustworthiness is not merely a model-level property. It is also a financial stability condition. Model failures, biased decision systems, privacy leakage, prompt injection, data poisoning, third-party concentration, algorithmic herding, and autonomous market responses can become systemic when many financial institutions rely on similar data, similar infrastructures, or similar foundation models. The paper proposes a framework that links trustworthy AI dimensions–robustness, safety, fairness, privacy, explainability, accountability, and human oversight– to core central bank functions. It concludes with a research agenda for central banks focused on model governance, AI-enabled supervision, privacy-preserving data cooperation, stress testing of AI-related shocks, and international coordination.

PDF