《“不役于物”——金融业新的AI监管期望》（英）-28页_367kb

Financial Stability Institute (FSI) Insights on Policy Implementation No 35

Humans Keeping AI in Check – Emerging Regulatory Expectations in the Financial Sector

Core Content

This paper explores the regulatory and supervisory expectations surrounding the use of artificial intelligence (AI) and machine learning (ML) in the financial sector. It highlights the growing recognition of AI's transformative potential in financial services, as well as the associated risks that require careful governance. The focus is on ensuring that AI is used responsibly, with an emphasis on fairness, transparency, accountability, and ethical considerations.

Main Viewpoints

• AI's Potential: AI and ML offer significant benefits in improving financial services delivery, operational efficiency, and risk management. These include better credit access, automated investment advice, efficient insurance claims processing, fraud detection, and risk scoring.
• Risks and Concerns: AI introduces new risks such as unintended bias, discrimination, lack of transparency, and potential exploitation of customers. These risks are not only relevant to conduct and consumer protection but also to prudential supervision.
• Regulatory Frameworks: Financial regulators are increasingly developing frameworks for AI governance, which often draw on existing principles of corporate governance, risk management, and ethical standards. These frameworks are still evolving and vary across jurisdictions.
• Key Principles: The paper identifies five core principles for AI governance: reliability/soundness, accountability, transparency, fairness, and ethics. These principles are similar to those applied to traditional models but are being adapted to address AI-specific challenges.
• Need for Human Oversight: There is a stronger emphasis on human involvement in AI systems to ensure accountability, fairness, and ethical use. Concepts like "human-in-the-loop" and "human-on-the-loop" are highlighted as important for model development and decision-making.
• Proportionality and Coordination: Regulatory and supervisory responses to AI must be proportional and coordinated, especially between prudential and conduct authorities. This is necessary to address the varying levels of risk and impact that AI models can have.

Key Information

• Regulatory Documents: The paper references policy documents from nine jurisdictions and international bodies, including the EU, France, Germany, Hong Kong, Luxembourg, the Netherlands, Singapore, the UK, and the US. These documents include principles, guidelines, and discussion papers.
• Common Principles:
  • Reliability/Soundness: Ensuring AI models are accurate and do not cause harm or discrimination.
  • Accountability: Clear roles and responsibilities, with a focus on human intervention and external accountability.
  • Transparency: Both internal and external transparency, including explainability, auditability, and disclosure to data subjects.
  • Fairness: Addressing and preventing biases in AI models to avoid discriminatory outcomes.
  • Ethics: Broader than fairness, covering societal norms, privacy, non-discrimination, and the ethical deployment of AI.
• Data Privacy and Protection: Regulators stress the importance of compliance with data privacy laws, especially as AI increasingly uses personal data. This includes ensuring data subject awareness, consent, and the ability to challenge AI-driven decisions.
• Third-Party Dependency: Financial institutions must manage risks related to third-party data and models, including lack of transparency, intellectual property constraints, and dependency risks.
• International Cooperation: The paper suggests that international standard-setting bodies may develop common guidance or standards for AI governance. This could help jurisdictions with emerging digital transformation efforts and provide a benchmark for AI deployment.

Challenges in Implementation

• Transparency: Ensuring AI systems are explainable and auditable, especially when dealing with "black box" algorithms.
• Reliability and Soundness: Assessing the accuracy and quality of AI models, particularly in the context of data bias and model errors.
• Accountability: Assigning clear responsibility for AI decisions, especially in cases where outcomes may be discriminatory or unethical.
• Fairness and Ethics: Defining and operationalising fairness in AI is complex, as it may not be explicitly addressed in consumer protection laws. There is also a need to balance "good biases" with potential discriminatory outcomes.
• Proportionality: Regulators must differentiate between AI models based on their risk and impact, applying more stringent oversight where necessary.

Conclusion

The use of AI in the financial sector is gaining momentum, and regulators are responding with evolving frameworks and guidance. While existing standards for traditional models apply to AI, there is a growing need for more specific, practical, and proportionate regulatory approaches. The paper concludes that international cooperation and the development of common standards could help address these challenges and support the responsible deployment of AI in financial services.

Annex – Proposed AI Regulation in the EU

The EU has proposed a regulation to harmonise AI rules across all industries, which includes principles on fairness, transparency, and accountability. This regulation is a pioneering effort in the global context and could serve as a model for other jurisdictions.

References

The paper references a range of regulatory documents and guidelines from financial authorities and international organizations, including the OECD AI Principles, G20 AI Principles, and various national regulatory bodies.