Model AI Governance Framework for Generative AI: A Singapore Guide

Singapore's Model AI Governance Framework for Generative AI, released by IMDA and the AI Verify Foundation on 30 May 2024, sets out nine dimensions for trustworthy generative AI, from accountability and data to content provenance and AI for public good. It is guidance, not legislation, and not a certification. Most of it applies to companies that deploy generative AI, not only to those that build models.

At a glance

Where the framework fits in Singapore's approach

Singapore has built its AI governance in layers. The original Model AI Governance Framework, first released on 23 January 2019 with a second edition on 21 January 2020, guides organisations deploying traditional AI. The generative AI framework expands it to the risks that come with large language and multimodal models, and IMDA has since published a framework for agentic AI, now at version 1.5 (May 2026).

Alongside the frameworks sit practical tools: IMDA's Starter Kit for testing LLM-based applications, and the AI Verify testing framework and Project Moonshot toolkit, explained in our guide to how AI Verify tests AI systems.

The law still applies on top. The PDPA governs personal data used in AI, and the PDPC has issued advisory guidelines on personal data in AI recommendation and decision systems (March 2024) and, on 20 July 2026, advisory guidelines on personal data in generative AI. If you are building a customer chatbot, read our PDPA checklist for AI chatbots in Singapore too.

The nine dimensions, and what they ask of your company

Each dimension below gives the framework's point in a sentence or two, then what a company that uses generative AI, rather than one training foundation models, can do about it.

1. Accountability

The framework asks for responsibility to be allocated upfront along the AI development chain, borrowing the shared responsibility model from cloud computing: each party answers for what it controls. It also suggests safety nets such as indemnity and insurance.

For you: name an accountable owner for each generative AI use case. Read your vendor contracts for who is responsible for model behaviour, data handling and intellectual property claims, and what indemnity, if any, you get.

2. Data

Data quality matters, and so does clarity on how personal data laws apply to AI. The framework mentions privacy-enhancing technologies and a pragmatic approach to copyrighted material in training data.

For you: know exactly what goes into prompts, retrieval indexes and fine-tuning sets. Keep personal data out unless you have a PDPA basis for using it, and check that you have the right to use the documents you feed into retrieval.

3. Trusted Development and Deployment

The framework expects baseline safety practices, such as input and output filters and retrieval-augmented generation backed by risk assessment, plus "food label" style disclosure: data used, evaluation results, mitigations, known risks and limitations, and intended use.

For you: ask vendors for that disclosure before you buy. For your own internal apps, write a one-page label: what the tool is for, what it is not for, its known limits and who to contact.

4. Incident Reporting

Developers should run vulnerability reporting channels, including bug bounty programmes, and have processes to notify the public and authorities about serious incidents, in proportion to the risk.

For you: add AI failures to your existing incident process: harmful output, confidential data exposed through prompts, wrong answers that customers acted on. Give staff and customers a way to report them. If personal data is involved, the PDPA's breach notification obligation may also apply.

5. Testing and Assurance

The framework wants a third-party testing ecosystem, common benchmarks and standards over time, and accreditation of testers for independence and competence.

For you: test before deployment, using the risks in IMDA's Starter Kit for LLM-based applications: hallucination, bias in decisions, undesirable content, data leakage and adversarial prompts. For high-impact apps, consider an external tester; the AI Verify Foundation's AI Tester Accreditation Programme exists for exactly this choice.

6. Security

Security-by-design has to adapt to natural-language inputs and probabilistic behaviour. The framework points to new safeguards such as input filters and digital forensics for models, and to threat modelling with resources such as MITRE ATLAS.

For you: treat prompt injection as a real risk, give AI tools the least access they need, and log what they do. Our explainer on prompt injection covers the basics for business teams.

7. Content Provenance

People need signals about where content comes from. The framework discusses digital watermarking and cryptographic provenance such as C2PA, interoperable standards, and work with publishers and platforms.

For you: label AI-generated content in customer-facing material, keep provenance metadata intact rather than stripping it, and make sure staff know which of your tools add it.

8. Safety and Alignment Research and Development

This dimension calls for faster research into aligned models, evaluation of dangerous capabilities and interpretability, coordinated across national AI safety institutes, including Singapore's.

For you: mostly a job for labs and researchers. Your part is to track model updates and vendor safety notes, and to re-test your apps when the underlying model changes.

9. AI for Public Good

The framework wants wider access to AI, including support for SMEs and digital literacy, responsible public-sector adoption, workforce upskilling and more sustainable AI through energy-efficient compute and carbon tracking.

For you: train staff before you roll tools out, and choose the smallest model that does the job well, since it is usually cheaper and uses less energy.

A one-page action list

If your system takes actions: the agentic AI framework

Generative AI that only writes text is one thing; agents that send emails, update records or make payments are another. IMDA's Model AI Governance Framework for Agentic AI, version 1.5 dated 20 May 2026, adds four dimensions: assess and bound the risks upfront, make humans meaningfully accountable, implement technical controls and processes, and enable end-user responsibility.

In practice that means deciding whether a task suits an agent at all, limiting each agent's tools and permissions by design, requiring human approval for actions that are hard to undo, rolling out gradually with monitoring, and telling users what the agent can do and what remains their responsibility.

Common misreadings

• "It only matters to big tech." Seven of the nine dimensions turn directly into actions for companies that simply use generative AI.
• "Following it means we comply with the law." It does not replace the PDPA or your sector regulator's requirements.
• "We need to be AI Verify certified." AI Verify is a testing framework and toolkit, not a certificate.
• "It is finished." The framework says it will evolve, and the agentic AI framework already extends it.

How CertTulen Academy can help

CertTulen Academy runs live, online, MCT-led training with hands-on practice. Leaders who will own AI policy and adoption can take AB-731: Drive AI Transformation in Your Organization, a one-day course aligned to Microsoft's AI Transformation Leader certification. For staff who need a shared foundation in using AI responsibly, CT-CLA-FLU1: AI Fluency: Framework and Foundations is a one-day course. We also run private cohorts for teams. Contact us for a batch that suits your time zone.