Singapore governs AI through a stack of voluntary frameworks and testing tools rather than a single AI statute, so organisations operating there have to work out which documents apply to them and what evidence each one expects. The stack has four layers: the Model AI Governance Framework (second edition, 2020) for traditional AI, the Model AI Governance Framework for Generative AI (2024), the framework for agentic AI (2026), and the AI Verify testing framework and toolkits that turn the principles into checks. This guide explains each layer, how they fit together, and how to build one governance and testing programme that satisfies all of them. It is for AI governance leads, risk and compliance teams, and engineering leads deploying AI in or for Singapore. The agentic AI framework has its own deep guide, the Singapore Model AI Governance Framework for Agentic AI Guide; this page covers it only where it connects to the rest.
Key takeaways
- Voluntary, but not optional in practice. The 2020 framework calls itself voluntary and "living", but it is the reference point the regulator and large customers use; binding duties still come from laws such as the Personal Data Protection Act 2012.
- Four key areas anchor everything. Internal governance, the level of human involvement, operations management, and stakeholder communication from the 2020 framework remain the backbone; later frameworks add generative and agentic risks on top.
- Human involvement is a risk decision. The 2020 framework's probability-severity of harm matrix chooses between human-in-the-loop, human-over-the-loop and human-out-of-the-loop.
- AI Verify turns principles into evidence. The AI Verify testing framework covers 11 principles with process checks and technical tests, was updated on 29 May 2025 to cover generative AI, and is crosswalked to the NIST AI RMF and ISO/IEC 42001.
- Testing guidance is now concrete. IMDA's Starter Kit for Testing LLM-Based Applications (version 1.0, January 2026) sets out risks, test methods and a six-step process you can adopt directly.
The Singapore AI governance stack
Each document builds on the earlier ones rather than replacing them. The table summarises status as published by IMDA, PDPC and the AI Verify Foundation, checked on 9 November 2026.
| Document or tool | Publisher and date | Covers | Status |
|---|---|---|---|
| Model AI Governance Framework, second edition | PDPC, 21 January 2020 | Organisations deploying AI at scale; principles and four key areas | Voluntary, living document |
| ISAGO and Compendium of Use Cases | IMDA and PDPC with the World Economic Forum Centre for the Fourth Industrial Revolution | Self-assessment against the 2020 framework; worked examples | Companion guidance |
| Model AI Governance Framework for Generative AI | AI Verify Foundation and IMDA, final 30 May 2024 (consultation from 16 January 2024) | Nine dimensions for a trusted generative AI ecosystem | Practical suggestions and first steps; non-binding |
| Model AI Governance Framework for Agentic AI | IMDA, version 1.0 January 2026, version 1.5 May 2026 | Organisations deploying agents built in-house or bought | Voluntary, living document |
| AI Verify testing framework and toolkit | IMDA and PDPC, minimum viable product 25 May 2022; open-sourced 7 June 2023; updated 29 May 2025 | 11 principles, process checks and technical tests for traditional and generative AI | Voluntary testing framework |
| Starter Kit for Testing LLM-Based Applications for Safety and Reliability | IMDA, version 1.0, January 2026 | Risk identification, test methods and thresholds for LLM applications | Voluntary; to be refined in stages |
The 2020 Model Framework: principles and four key areas
The second edition rests on two guiding principles: AI-assisted decisions should be explainable, transparent and fair, and AI solutions should be human-centric, with people's interests, well-being and safety as primary considerations. It is aimed at organisations deploying AI at scale and does not cover off-the-shelf software that merely includes AI features. It turns the principles into four key areas.
- Internal governance structures and measures: adapt existing structures or create new ones, with clear roles and responsibilities for AI and risk management and internal controls around it. In practice: an AI register, a named accountable owner per system and an approval gate before deployment. See the AI Governance Guide.
- Determining the level of human involvement: use the probability and severity of harm, plus the nature and reversibility of harm and operational feasibility, to choose one of three approaches (below).
- Operations management: data lineage, quality and bias mitigation, dataset review, and model measures such as explainability, repeatability, robustness, regular tuning, traceability, reproducibility and auditability.
- Stakeholder interaction and communication: disclose AI use, set a policy on explanations, offer opt-out where appropriate, provide feedback and review channels, test user interfaces, and communicate with other organisations in the supply chain.
Choosing human involvement
| Approach | What it means in the framework | When it fits (NHI Mgmt Group reading) | Control to evidence |
|---|---|---|---|
| Human-in-the-loop | A human makes the final decision; the AI recommends | High severity, high probability of harm; irreversible outcomes such as credit refusal or account termination | Decision records showing human approval, override rates |
| Human-over-the-loop | A human supervises and can take over or adjust | Moderate severity, or harms that can be reversed quickly | Monitoring dashboards, alert thresholds, intervention logs |
| Human-out-of-the-loop | No human oversight of execution | Low severity and low probability, high volume decisions such as product recommendations | Periodic outcome review, drift monitoring |
For AI agents, this choice becomes an access decision: an out-of-the-loop agent needs tightly scoped credentials because nobody approves each action. The agentic framework develops this further; see the Singapore Model AI Governance Framework for Agentic AI Guide.
The generative AI framework: nine dimensions
The 2024 framework, developed by the AI Verify Foundation and IMDA, widens the view from a single deploying organisation to the whole ecosystem of model developers, application deployers and users. Its nine dimensions, as summarised on the AI Verify Foundation's framework page:
| Dimension | Focus | What a deployer can do |
|---|---|---|
| Accountability | Align incentives so that parties across the development life cycle are responsible to end users | Allocate responsibility in contracts with model and platform providers |
| Data | Data quality and pragmatic handling of contentious training data | Record data sources for fine-tuning and retrieval; apply PDPA rules |
| Trusted development and deployment | Transparency about baseline safety and hygiene practices | Ask providers for model cards and evaluation results; publish your own disclosures |
| Incident reporting | Prompt reporting, remediation and learning | Add AI-specific categories to the incident process. See the AI Incident Reporting Guide |
| Testing and assurance | Third-party testing and common testing standards | Use AI Verify, Project Moonshot and the Starter Kit |
| Security | New threats introduced by generative models | Threat model prompt injection, data leakage and tool misuse |
| Content provenance | Signals to users about where content comes from | Label AI-generated content; consider content credentials. See the AI Content Provenance and C2PA Guide |
| Safety and alignment R&D | Global cooperation among AI safety institutes | Track published research and evaluation methods |
| AI for public good | Access, public sector adoption, upskilling, sustainability | Mostly for government and industry bodies |
AI Verify and the testing toolkits
AI Verify is Singapore's AI governance testing framework. The AI Verify Foundation's description lists 11 principles: transparency, explainability, reproducibility, safety, security, robustness, fairness, data governance, accountability, human agency and oversight, and inclusive growth, societal and environmental well-being. Each principle is broken into outcomes, processes and evidence; processes are validated by documentary evidence. The framework is delivered as a PDF overview, an Excel tracker for process checklists and a Process Checks tool for generative AI that produces a summary report.
- Technical testing: the AI Verify Testing Toolkit covers traditional AI; Project Moonshot provides technical testing for generative AI. The framework is designed to combine technical tests with process checks where applicable.
- Crosswalks: the Foundation publishes crosswalks to the NIST AI RMF (13 October 2023), the NIST AI RMF Generative AI Profile (29 May 2025), the Hiroshima Process International Code of Conduct (29 May 2025) and ISO/IEC 42001:2023 (21 July 2025). If you already run one of these, start from the crosswalk rather than a fresh assessment. See the NIST AI RMF Guide and the ISO/IEC 42001 Guide.
- Global AI Assurance Pilot: IMDA and the AI Verify Foundation ran a pilot from March to May 2025 in which specialist testers assessed generative AI applications from 17 organisations across 10 sectors. Findings published on 29 May 2025 stressed that human experts are needed at every stage of testing, and the pilot led to a voluntary testing starter kit.
The Starter Kit for LLM applications
Version 1.0 of IMDA's Starter Kit (January 2026) is the most directly usable document in the stack for engineering teams. It covers five risks: hallucination and inaccuracy (including grounding in source documents), bias in decision making, undesirable content, data leakage, and vulnerability to adversarial prompts, both direct and indirect. It describes four main methods: output testing of the whole application, component testing of filters, system prompt, retrieval and model, benchmarking against reference answers, and red teaming, with user simulation, early-user pilots and A/B testing as supplements. Its process runs: identify material risks and set thresholds before testing (with hard gates such as zero tolerance for illegal content), test outputs first and components when results fall short, assess and mitigate then re-test, check consistency through repeated runs and validate any LLM-as-a-judge evaluator with human experts, monitor after deployment, and look for newer benchmarks. See the AI Red Teaming Guide.
Building one programme across the frameworks
- Inventory and classify. List every AI system, generative application and agent, with owner, purpose, data used, provider and whether it makes or supports decisions about individuals.
- Set human involvement per system. Apply the 2020 probability-severity matrix and record the chosen approach and rationale. For agents, record the actions each agent may take without approval.
- Map duties once. Use the AI Verify process checklist as the master control list, mapped to the generative AI dimensions and, through the published crosswalks, to NIST AI RMF or ISO/IEC 42001 if you also report against them.
- Test before release. Apply the Starter Kit process to each LLM application: thresholds first, output tests, component tests where needed, red teaming for high-risk uses. Run it in a pre-production environment and keep the reports.
- Disclose and listen. Publish AI use notices, explanation policy and a feedback route, and connect complaints to the incident process.
- Monitor and re-test. Re-run tests after model, prompt or retrieval changes. Roll back to the previous model or prompt version if a threshold fails in production.
A small organisation using one vendor chatbot can meet the expectations with a one-page register entry, a human-over-the-loop decision, vendor evidence on testing and a short output test against the Starter Kit risks. A large organisation with many AI systems should run AI Verify process checks per system, keep a central testing team, and report coverage to its risk committee.
Where identity fits
The frameworks speak of accountability and human oversight, but in production these depend on identity. Each AI system and agent needs its own non-human identity, with credentials scoped to its approved actions, so that logs show which system acted and oversight can be enforced rather than assumed. The Starter Kit's data leakage risk is often an access control problem: a retrieval pipeline that can read documents its users cannot. Apply least privilege to the retrieval service account, not only the prompt. See the Agentic AI Identity Guide. Personal data used in AI remains subject to the PDPA; see the Singapore PDPA Guide.
Detection and measurement
- Register coverage: percentage of AI systems and agents in the register with an owner and a recorded human involvement decision.
- Pre-release testing: percentage of LLM applications tested against all five Starter Kit risks before release.
- Threshold breaches: number of production test runs that fail a set threshold, and time to fix.
- Process check completion: AI Verify process checks completed per high-risk system.
- Human override rate: for human-in-the-loop systems, how often reviewers change the AI recommendation (a rate near zero may mean rubber-stamping).
- AI incidents: incidents and user complaints attributed to AI systems, by category.
Common failure patterns
- Treating voluntary as irrelevant. The frameworks set the expectations customers, partners and the regulator use; ignoring them leaves no defensible position when something goes wrong.
- One-off testing. A system tested at launch and never again drifts as models, prompts and retrieval data change.
- Human-in-the-loop in name only. Reviewers approve every recommendation without the time or information to challenge it.
- Testing the model, not the application. The Starter Kit tests the whole application, including filters, prompts and retrieval; a model benchmark alone misses most failures.
- Shared credentials for AI services. Several agents or pipelines using one API key make accountability and containment impossible.
Framework mapping
| Singapore element | NIST AI RMF function | ISO/IEC 42001 area |
|---|---|---|
| Internal governance structures and measures | Govern | Leadership, roles and AI policy |
| Human involvement (probability-severity matrix) | Map | AI risk assessment and impact assessment |
| Operations management; AI Verify technical tests | Measure | Operation, data and system life cycle controls |
| Incident reporting; monitoring and re-testing | Manage | Performance evaluation and improvement |
This is a high-level NHI Mgmt Group mapping; use the AI Verify Foundation's published crosswalks for control-level detail.
Practitioner checklist
- Record which Singapore frameworks your organisation commits to, and the version and date checked.
- Keep an AI register covering traditional AI, generative applications and agents.
- Record a human involvement decision and rationale for every system that affects individuals.
- Adopt the AI Verify process checklist as the master control list, mapped through the crosswalks.
- Test every LLM application against the five Starter Kit risks with thresholds set in advance.
- Validate any LLM-as-a-judge evaluator with human experts before relying on it.
- Give each AI system and agent its own scoped non-human identity.
- Publish AI use notices and a feedback route, and connect complaints to incident handling.
Standards and references
- PDPC: Model Artificial Intelligence Governance Framework, Second Edition
- AI Verify Foundation and IMDA: Model AI Governance Framework for Generative AI
- AI Verify Foundation: What is AI Verify
- IMDA: Starter Kit for Testing LLM-Based Applications for Safety and Reliability
- Singapore EDB: Initiative by IMDA and AI Verify Foundation Tests AI Accuracy and Trustworthiness
- NIST: AI Risk Management Framework
Related NHI Mgmt Group resources: Singapore Model AI Governance Framework for Agentic AI Guide · Global AI Regulation Map · Singapore PDPA Guide · AI Governance Guide. This guide summarises Singapore's voluntary AI governance frameworks and is not legal advice.