What is AI regulation in Singapore?
Global AI regulation
Singapore regulates AI through a governance-first model, not through a single omnibus AI Act. In practice, that means voluntary but influential frameworks from IMDA and PDPC, especially the Model AI Governance Framework, AI Verify, and the 2026 agentic AI guidance, sit alongside binding laws such as the PDPA and other existing rules on online harms, contracts, torts and sector supervision. For organisations, the real job is to show accountability, testing, human oversight, data protection and clear user communication.
What this means
In Singapore, "AI regulation" does not mean only statutes. It includes hard law, official guidance, testing frameworks and assurance practices. The state has deliberately built a practical architecture that tells organisations how to govern AI responsibly, then uses existing laws where personal data, safety, misleading conduct, online harms or sector-specific duties are involved.
That makes Singapore different from jurisdictions that started with a single AI statute. The Singapore approach began with a general model framework for AI use at scale, then expanded for generative AI and later for agentic AI. AI Verify and related testing tools translate governance ideas into auditable evidence, so organisations can do more than make broad claims about responsible use.
Why it matters
Organisations sometimes misread Singapore because they look only for a headline AI law and miss the rest of the stack. That is risky. If you deploy AI in Singapore, especially with personal data, customer-facing recommendations, automated decision support, or agents that can call tools and act on other systems, you still need named owners, written policies, security controls, testing records, vendor controls and incident processes.
This matters beyond enforcement. Boards, procurement teams, enterprise buyers and regulated customers increasingly want proof that AI has been governed sensibly. Singapore's model is designed to create that proof. It gives organisations a shared vocabulary for accountability, a path to technical testing and a way to align internal practice with internationally recognised governance expectations without waiting for a single new Act.
How it works
Singapore uses a governance-first stack
As of June 2026, Singapore's AI regime is best understood as a layered stack. At one layer are voluntary governance frameworks and testing mechanisms. At another are binding laws that already apply to the facts of a deployment. So when people ask "what is AI regulation in Singapore?", the answer is not one document. It is a combination of official governance frameworks, assurance tools and ordinary law.
That distinction matters. Governance frameworks do not themselves create a licensing regime or a general pre-market approval system for AI. But they strongly shape what responsible practice looks like in Singapore. Existing law, especially data protection law, still does the binding work when personal data is collected, used, disclosed, retained, transferred or exposed, and when unsafe or harmful uses of technology trigger other legal regimes.
The 2020 model framework remains the baseline
Singapore's baseline architecture is the Model Artificial Intelligence Governance Framework, with the second edition released in January 2020. It is a voluntary baseline that is meant to evolve over time. It is also deliberately broad. The framework is algorithm-agnostic, technology-agnostic and sector-agnostic, so it can be applied across different use cases and industries.
The 2020 framework is built around two high-level ideas. First, AI-assisted decision-making should be as explainable, transparent and fair as is reasonable in context. Second, AI should remain people-centred. Operationally, the framework focuses on four areas: internal governance structures and measures, the level of human involvement in AI-augmented decision-making, operations management, and stakeholder interaction and communication.
That is important because Singapore does not treat responsible AI as only a technical matter. The 2020 framework asks organisations to appoint clear internal owners, fit AI into enterprise risk management, decide where human review is needed, manage data and model operations through the lifecycle, and communicate with affected stakeholders in a way that builds trust. It also makes clear that following the framework does not excuse compliance with existing law. Instead, it helps demonstrate accountability-based practice.
The baseline framework also has a scope boundary that is often missed. It is mainly aimed at organisations deploying AI at scale in products, services or operations. It is not written chiefly for buyers of an ordinary off-the-shelf software package that happens to contain an AI feature. Even so, that does not mean those buyers are free from legal duties. Their actual use of the tool can still trigger PDPA, security, procurement and disclosure questions.
Generative AI expanded the model into an ecosystem view
In 2024, Singapore extended the architecture with the Model AI Governance Framework for Generative AI. This did not discard the earlier framework. It expanded it. The GenAI framework moves from a narrower organisational checklist to a broader ecosystem view, recognising that generative AI risk is spread across developers, deployers, hosts, researchers, policymakers and end-users.
The GenAI framework is organised around nine dimensions: accountability, data, trusted development and deployment, incident reporting, testing and assurance, security, content provenance, safety and alignment research and development, and AI for public good. That matters for two reasons.
First, it shows that Singapore does not see generative AI governance as only a question of model behaviour. It also treats disclosure, data quality, incident practice, provenance signals and third-party testing as part of the same architecture. Second, it keeps the regime durable. Even if specific models and threat types change, the governing dimensions remain recognisable.
The GenAI framework also presents itself as an initial step rather than a finished code. In other words, it is meant to be extended through implementation guidance, testing practice and further resources, not treated as a closed rulebook.
Agentic AI guidance focuses on autonomy, permissions and control
In January 2026, IMDA issued the first Model AI Governance Framework for Agentic AI, and an updated version followed in May 2026 after feedback from more than 60 companies. This is one of the clearest signs of Singapore's regulatory style. Rather than rushing straight to a new statute, it published more detailed governance guidance for the next class of systems as soon as practical experience started to surface.
The agentic framework is useful because it defines the problem more precisely than generic "GenAI" language. It focuses on agents built on generative AI models and tells organisations to think in terms of two variables: action-space and autonomy. Action-space is the range of actions the agent can actually take, including what tools it can call, which systems it can access, and whether it can read only or also write, transact or change an environment. Autonomy is the degree to which the agent decides how to act towards a goal.
Once you frame agentic AI that way, the governance task becomes clearer. Singapore's guidance is arranged around four dimensions: assess and bound the risks upfront; make humans meaningfully accountable; implement technical controls and processes; and enable end-user responsibility. The framework stresses that human responsibility does not disappear because an agent can act by itself. Responsibility has to be allocated clearly across product, security, business and user roles, and human oversight must remain effective over time rather than becoming a box-ticking ritual.
The agentic guidance is also more operational than the earlier material in some respects. It pushes organisations towards bounded permissions, narrow tool access, phased deployment, continuous logging and monitoring, change management, post-deployment testing and user training. It is especially concerned with what happens when agents are connected to internal systems, third-party tools, browsers, files, payments or other high-impact actions.
Binding duties come mainly from the PDPA and existing law
For most private organisations, the first hard-law contact point is still the Personal Data Protection Act 2012. The PDPA requires organisations to designate one or more individuals as responsible for compliance, maintain policies and practices, protect personal data with reasonable security arrangements, and assess and notify certain notifiable breaches. The PDPC can issue directions and financial penalties for non-compliance.
For AI, the PDPA matters earlier in the lifecycle than many teams expect. Singapore's 2024 Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems make that explicit. They explain how the PDPA applies not only at deployment, but also during development, testing and monitoring, and during procurement of bespoke AI systems.
Those guidelines are operationally important. They explain when meaningful consent is needed and when statutory exceptions may be relied on, especially the business improvement and research exceptions. They encourage data minimisation, anonymisation where feasible, impact assessments where useful, and written policies explaining practices and safeguards. They also connect AI governance to transparency. In higher-impact contexts, organisations are encouraged to explain how fairness, reasonableness, oversight, safety and robustness have been addressed.
This part of the regime also matters for vendors and integrators. Where a service provider develops a bespoke or fully customisable AI system using the client's personal data, it can become a data intermediary under the PDPA. That means real protection and retention duties, plus practical expectations around data mapping, provenance records and support for the client's own consent, notification and accountability duties.
Beyond data protection, Singapore has confirmed that existing legislation and ordinary legal principles still apply to AI-related harm. The Government has pointed specifically to online harms and misinformation rules, as well as contract and tort law, while also saying it is still studying whether accountability gaps could justify more targeted policy, regulatory or legal measures in future.
AI Verify turns governance claims into evidence
AI Verify is central to understanding Singapore's model because it sits between governance and assurance. It is not a statute, a licence or a legal safe harbour. It is a testing framework that helps organisations assess the responsible implementation of AI against 11 principles, including transparency, explainability, repeatability and reproducibility, safety, security, robustness, fairness, data governance, accountability, human agency and oversight, and broader societal and environmental well-being.
What makes AI Verify distinctive is the evidence model. It does not stop at high-level principle statements. It links principles to concrete processes and documentary evidence. Organisations can record process checks and generate a summary report showing how they have implemented the framework. That is useful for internal compliance teams, external reviewers and counterparties who want more than policy slogans.
AI Verify has also evolved with the technology. The framework was updated in May 2025 so it can be used for both traditional AI and generative AI applications. It is mapped to other international instruments, including the US NIST AI RMF, the GenAI profile, the Hiroshima Process code of conduct and ISO/IEC 42001. That crosswalk matters for companies working across jurisdictions, because it reduces the need to treat Singapore governance work as a dead-end local exercise.
Testing and assurance are part of the architecture, not an afterthought
Singapore has pushed strongly on practical testing. In 2024, IMDA launched Project Moonshot, an open-source toolkit intended to bring benchmarking, red teaming and baseline testing for large language model applications into one usable platform. IMDA then followed with more detailed testing guidance for LLM-based applications, positioning that material as a technical companion to the GenAI framework's testing and assurance dimension and to AI Verify's process checks.
The practical message is straightforward. In Singapore's model, responsible AI is not only policy writing. It is pre-deployment testing, clear thresholds for risk tolerance, and post-deployment monitoring. This becomes even more important for agentic systems because the environment keeps changing around them and small alterations can have cascading effects across multiple tools or agents.
So the mechanics of compliance and governance in Singapore increasingly look like this: define the use case and risk profile, decide which framework layer applies, assign human responsibility, document data use, test the system in a structured way, disclose the right information to users and stakeholders, and keep monitoring once it is live.
Examples
A useful disclosure example comes from the PDPC's own materials. In RE HSBC [2021] SGPDPC 3, cited in the 2024 AI advisory guidelines, HSBC was treated as having met accountability and disclosure duties by providing information about how it used personal data and AI technology in credit facility assessments. The practical lesson is that in Singapore, governance is not only internal. Public-facing explanations can matter.
A useful agentic example appears in IMDA's 2026 framework through OCBC. Bank of Singapore relationship managers and compliance staff use an agentic system that parses financial documents and drafts a source of wealth memo. The system is framed as decision support only. It is meant to improve consistency and reduce manual effort, but final validation and approval remain with human staff. That matches Singapore's emphasis on bounded autonomy and meaningful human accountability.
A useful assurance example comes from the Global AI Assurance Sandbox case study involving CDL and Knovel. CDL's internal agentic system could retrieve internal knowledge and perform multi-step workflows. The priority risk selected for testing was data leakage across users. The assurance work focused on whether access controls and boundaries were actually enforced, in other words whether a user could obtain only the information they were entitled to access. That is a strong example of how Singapore links governance language to concrete technical testing.
Common misunderstandings
"Singapore has no AI regulation because it has no AI Act." This is wrong. Singapore's model combines official governance frameworks with binding laws that already apply.
"The Model AI Governance Framework is mandatory law." It is not. It is voluntary guidance, but it is influential and often treated as a practical benchmark.
"AI Verify is a certification or a legal shield." It is neither. It is a testing and documentation framework that helps create evidence.
"The PDPA matters only after launch." It does not. In Singapore's own guidance, PDPA questions arise during development, testing, monitoring, deployment and bespoke procurement.
"If an agent acts autonomously, the human owner is no longer responsible." The 2026 guidance says the opposite. Human responsibility still applies and must be allocated clearly.
Risks and boundaries
Singapore's model has real strengths, but it has boundaries. The frameworks are not statutes, do not create automatic immunity and do not substitute for sector-specific or fact-specific legal analysis. AI Verify reports can support governance claims, but they are not government approval and do not by themselves prove lawful deployment.
The model also remains evolutionary. The GenAI and agentic frameworks are explicitly living documents. The Government has said existing legislation and ordinary private law continue to apply, while it studies whether accountability gaps may need further policy, regulatory or legal measures. So the durable point is this: Singapore currently expects organisations to govern AI responsibly within an existing legal framework, while remaining ready for targeted hard-law changes if specific risks demand them.
What to do next
Start with an inventory. Separate ordinary software features from AI systems you genuinely configure, deploy or rely on for business processes. Then classify those systems by personal data use, degree of autonomy, external-system access, possible harm, and whether they support or replace human judgment in any material step.
Next, map each use case to the right Singapore layer. Use the 2020 framework as the baseline. Add the GenAI framework if the system generates content or uses foundation models. Add the agentic framework if it can plan, call tools, browse, write, transact or operate through multi-step workflows. Tie that governance map to PDPA controls, named internal accountability, vendor terms, security controls, disclosure practice, incident handling and testing evidence.
Finally, insist on proof. For higher-risk systems, do not rely on policy decks alone. Use structured testing, documentary evidence, logs, version control, phased rollout and user training. For agentic deployments, begin with narrow permissions and clear approval gates, especially where agents can reach sensitive data or change live systems.
FAQs
Does Singapore have a single AI Act like the EU AI Act?
Not currently. Singapore's present model is built from voluntary governance frameworks and binding existing laws, especially the PDPA and other applicable legal regimes.
Is the Model AI Governance Framework mandatory?
No. The framework family is voluntary guidance. It matters because it provides the official benchmark for responsible AI practice in Singapore's governance-first model.
Is AI Verify mandatory?
No. AI Verify is a testing and documentation framework. It helps organisations assess their practices and generate evidence, but it is not a statutory requirement in itself.
When does Singapore's data protection law become relevant to AI?
As soon as personal data is collected, used or disclosed in development, testing, monitoring, deployment or bespoke procurement. That can trigger duties around consent analysis, exceptions, security, policies, DPO responsibility and breach handling.
What changes when a system becomes "agentic"?
The governance focus becomes sharper. Singapore expects tighter control over tool permissions, data access, write actions, approvals, logging, monitoring, change management and user training, because agents can affect the world more directly than a simple chatbot.
Do bespoke AI vendors have their own duties under Singapore law?
Often yes. If they process a client's personal data while building a bespoke or fully customisable system, they can be treated as data intermediaries under the PDPA, with protection and retention duties and a practical role in supporting the client's own compliance.
If we only buy standard AI software, can we ignore the frameworks?
No. The 2020 baseline is mainly aimed at organisations deploying AI at scale, but your actual use of the software can still raise PDPA, security, procurement and disclosure issues. You still need proportionate governance.
Is Singapore likely to add more AI-specific law later?
Possibly, but nothing general has been confirmed. The Government has said it is still studying whether accountability gaps justify further policy, regulatory or legal measures.