HKMA and Cyberport Launch Second Cohort of GenAI Sandbox to Drive Banks’ AI Adoption

The Hong Kong Monetary Authority (HKMA) and Hong Kong Cyberport Management Company Limited (Cyberport) have launched the second cohort of the Generative Artificial Intelligence (GenAI) Sandbox initiative to provide a risk-controlled environment for banks to develop and test AI-powered solutions. In its press release of 28 April 2025, the HKMA encourages banks to participate in the second cohort of the GenAI Sandbox to stay ahead in AI-driven risk management and, in particular, to proactively defend themselves and their customers against increasingly prolific deepfake scams.

KEY DEVELOPMENTS

1. Expansion of the GenAI Sandbox Initiative

Building on the success of the first cohort launched in January 2024, the second cohort of the HKMA’s GenAI Sandbox will continue to explore high-impact use cases, focusing in particular on:

• Risk management enhancements;
• Anti-fraud measures; and
• Improvements in customer experience.

2. Introduction of the GenAI Sandbox Collaboratory Platform

A notable addition to the second cohort is the GenAI Sandbox Collaboratory platform, a structured platform hosting practical workshops designed to:

• Facilitate early engagement between banks and technology providers;
• Translate problem statements into actionable AI use cases; and
• Accelerate the development of regulation-compliant solutions.

A dedicated workshop on combating deepfake scams will also be held in the coming weeks, bringing together tech firms, digital banks and SMEs to share detection tools and mitigation strategies against AI-enabled fraud.

3. Strategic Focus on “AI vs. AI” and Risk Mitigation

According to HKMA Deputy Chief Executive, Mr. Arthur Yuen, this second cohort will explore the potential of “AI versus AI”, that is leveraging AI to counter emerging AI-related risks. The HKMA encourages banks to integrate AI into their second and third lines of defence to strengthen risk management frameworks while benefiting from AI developments.

4. FiNETech5 Event Highlights Industry Collaboration

The launch of the second cohort was announced by Carmen Chu, HKMA Executive Director (Banking Supervision) (Ms Chu), during FiNETech5, an event attended by over 150 AI practitioners from the banking and tech sectors. Key takeaways from her Opening Remarks at FiNETech5 covered the following areas.

Data Readiness: The Foundation for Effective GenAI Deployment

• Ms Chu emphasised the need for financial institutions to invest in scalable data infrastructure and governance frameworks to support their AI adoption;
• Commenting that high-quality data is essential for training and testing AI models, and that data collection and annotation can be time-intensive, particularly for unstructured or novel data formats, she noted the need for banks to establish repeatable pipelines to process unstructured data efficiently (e.g., fraud investigations involving financial proofs, application forms and transaction histories); and
• Some banks, for example, are leveraging GenAI to synthesise and preprocess data, allowing them to accelerate preparation while ensuring accuracy.

Holistic Model Evaluation: Beyond Conventional Metrics

• Ms Chu noted the need for financial institutions to adopt a holistic approach to evaluating model performance. While public benchmarks can provide a useful reference point, banks need to develop their own custom evaluation metrics tailored to their use cases.
• For customer-facing applications (e.g., GenAI chatbots), continuous post-deployment monitoring is also critical to ensuring accuracy, professionalism and reliability.
• Unlike traditional User Acceptance Testing—which uses binary pass/fail criteria—GenAI solutions require multidimensional metrics (e.g., response completeness, contextual accuracy and consistency over time).

Rethinking Fine-Tuning Strategies for Large Language Models (LLMs)

• Ms Chu also made the point that banks should critically reassess if legacy fine-tuning approaches are cost-effective or if newer LLMs can achieve comparable results with minimal customisation.
• Early adopters fine-tuned models extensively (e.g., using past samples to structure outputs such as funding proposals).
• However, advancements in LLMs—such as improved reasoning and “mixture of experts” capabilities—raise questions as to whether heavy fine-tuning is still necessary.

Conclusion: Strategic Priorities for Banks

The key takeaways for banks are that they need to:

• Invest in data infrastructure to handle unstructured data at scale;
• Develop tailored evaluation frameworks to ensure GenAI solutions meet operational and regulatory standards; and
• Re-evaluate fine-tuning practices in light of evolving LLM capabilities to optimise resource allocation.