What does SFC Circular 24EC55 (generative AI language models) say about AI in banking?

What are the key points of SFC Circular 24EC55 (generative AI language models)?

• Core Principle 1 (senior management responsibilities): effective policies, procedures and internal controls and suitably qualified oversight across the full AI LM lifecycle, including identifying high-risk use cases and having staff with AI, data science, model risk and domain competence.
• Core Principle 2 (AI model risk management): segregate model development from validation where practicable; validate before approval and on material change; test end-to-end including retrieval-augmented generation (RAG) and content filtering; monitor ongoing performance.
• Paragraph 8: using an AI LM for investment recommendations, advice or research is generally a high-risk use case (this excludes after-sales client servicing, per footnote 3).
• Paragraph 17: where an AI LM is used in the client interface, prominently disclose that the client is interacting with AI and that output may be inaccurate. Paragraphs 18-19 add, for high-risk uses, validation, a human in the loop to review factual accuracy before output reaches the user, output-robustness testing against prompt variations and disclosure at each interaction.
• Core Principle 3 (cybersecurity and data risk management): cover adversarial attacks on AI LMs and their training data, test adversarially to the extent practicable, encrypt non-public data, control confidential data input and mitigate browser-extension leakage; LCs should have regard to the Privacy Commissioner's Model Personal Data Protection Framework for AI.
• Core Principle 4 (third-party provider risk management): due diligence and monitoring of providers, assessing the provider's model risk framework where transparency is limited, supply-chain and data leakage assessment of components such as embedding models and vector stores, and contingency plans for disruption.
• Paragraph 30: LCs intending to adopt AI LMs in high-risk use cases are reminded to comply with the notification requirements under the Securities and Futures (Licensing and Registration) (Information) Rules and are encouraged to discuss plans with the SFC early.
• The circular takes immediate effect; the SFC says it recognises some LCs need time to update policies and will take a pragmatic approach in assessing compliance, and a risk-based implementation is allowed.

What did SFC Circular 24EC55 (generative AI language models) change for banks?

It was the first SFC statement of expectations on generative AI. It ties AI LM use to existing obligations (the Code of Conduct and the Management, Supervision and Internal Control Guidelines) and introduces the high-risk use case concept for investment recommendations, advice and research, with human-in-the-loop and notification expectations. The SFC's 2 June 2026 circular on AI-enabled cyberattacks builds on its cybersecurity principles.

What does SFC Circular 24EC55 require of firms using generative AI language models?

SFC Circular 24EC55 of 12 November 2024 requires licensed corporations that use generative AI language models in regulated activities to meet four Core Principles: senior management must put effective policies, controls and qualified oversight in place across the AI LM lifecycle; AI model risk management must include segregated development and validation, validation before use and on material change, end-to-end testing and ongoing monitoring; cybersecurity and data risk management must cover adversarial attacks, encryption and confidentiality of client data; and third-party provider risks, including supply chain and concentration, must be assessed with contingency plans. High-risk uses, meaning investment recommendations, advice or research, need extra measures including a human in the loop, prompt-robustness testing and disclosure at each client interaction, and firms must observe notification requirements when adopting them. It applies from 12 November 2024 and is addressed to licensed corporations rather than banks.

The circular is risk-based: LCs may implement the Core Principles in a manner commensurate with the materiality of impact and the level of risk of each use case. Model Development requirements apply only where an LC develops, customises, refines or enhances an AI LM (for example fine-tuning, RAG, content filtering or integrating prompt-management tools); an off-the-shelf product used with only essential parameters configured remains subject to Model Management but not Model Development requirements.

It anchors each expectation to existing SFC instruments, citing the Code of Conduct and the Management, Supervision and Internal Control Guidelines in its footnotes, and footnote 13 states that the AI model risk framework need not duplicate firms' existing cybersecurity, data and third-party frameworks provided those cover the circular's requirements. Hong Kong's banks fall under the HKMA, whose guidance sits outside this circular.

WHAT THIS MEANS IN PRACTICE

• Inventory every AI LM use in regulated activities and classify which, if any, are high-risk investment recommendation, advice or research uses.
• Segregate development and validation functions where practicable and document validation, end-to-end tests and monitoring.
• Add AI disclosures and, for high-risk uses, human review before output reaches clients.
• Extend third-party risk management to AI providers, embedding models, vector stores and orchestration tools.
• Check whether adopting a high-risk use triggers notification under the Information Rules and speak to the SFC early.
• Review prompt-robustness and adversarial testing practices.