AI Model Lifecycle Governance and Human Oversight
Tags: Frameworks
Lifecycle Stages
The Guideline structures risks and responsibilities across four stages:
- Planning & Data Collection: Risks include harmful data/poisoning, data bias, personal data privacy, and IP concerns. Calls for diversified sources, distribution analysis, filtering harmful content, and proper authorisation. Notes value of domain-specific fine-tuning (e.g., legal, regulatory, government, banking) for accuracy and compliance.
- Model Development: Addresses overfitting/underfitting and IP risks when using third-party architectures or pre-trained weights; emphasises metrics, evaluation, and licence compliance.
- Deployment & Integration: Identifies system integration and access control risks; advises redundancy, comprehensive testing, and strict permissioning. Discusses deployment choices (cloud, on-premise open-source, on-device) and hybrid strategies based on data criticality. Encourages interoperability standards (open APIs, secure data-sharing) and regulatory alignment with recognised frameworks.
- Usage & Maintenance: Highlights content safety, data exposure (including via RAG), copyright complexities, and credibility/ethical/social risks. Recommends input filtering, content labelling, clear documentation, and user verification of outputs.
Human Oversight
- Collaborative Generative AI Models: Lower-impact scenarios where human judgment complements AI with limited oversight needs.
- Human-Dominated Models: Higher-impact contexts relying primarily on human decision-making with AI as an auxiliary tool.
Public Service Applications
The Guideline notes that multimodal voice assistants can augment government service channels (e.g., speech recognition, sentiment analysis, voice-to-text summarisation), handling routine inquiries and informing policy through insights while reserving complex cases for human agents.
Overall, the lifecycle view ties concrete safeguards to each stage while calibrating human oversight to potential impact, reinforcing accountability and trust.
Sources:- HK_Generative_AI_Technical_and_Application_Guideline_en.pdf
← Back to Knowledge Base