Singapore AI Governance in 2026: Model Frameworks, AI Verify, PDPA

Quick verdict

Singapore still governs AI with voluntary Model Frameworks plus existing law. IMDA’s Model AI Governance Framework (2019/2020), the generative-AI framework (May 2024), and the Model AI Governance Framework for Agentic AI (22 January 2026, updated 20 May 2026) tell you how to run systems. AI Verify is how you test the claim. The Personal Data Protection Act is how you get fined if personal data in a prompt, log or vendor model is mishandled.

If you already operate in Australia, do not build a second religion. Build one evidence layer: inventory, data, vendor, owner, risk, testing, human control. Map it to IMDA language for Singapore and to VAISS / APP 1 for Australia.

Best for: Singapore companies and AU/SG groups that need one register. Honest limit: MAS, MOH and other sector regulators can impose binding technology duties we do not replace. We are not a PDPC adviser of record.

Last updated: 31 August 2026.

What AI governance obligations does a Singapore company have in 2026?

Binding: PDPA (and sector rules if you are in them). Expected by counterparties: alignment with the Model Frameworks and, for higher-risk or customer-facing systems, an AI Verify-style test pack. There is no Singapore statute that says “implement the Model Framework or you are unlawful.” Organisations remain accountable for what their agents do under ordinary law.

The named person for the binding half is the Data Protection Officer. Section 11 of the PDPA requires every organisation to designate one and publish their business contact. If you are selling or buying an AI evidence system in Singapore, that is who sits in the room.

The three Model Frameworks, in one table

Framework	When	What it adds	Evidence we keep
Model AI Governance Framework	2019, 2nd ed. 2020	Internal governance, human oversight, risk, user communication	Owner, risk tier, oversight design, user notice
Model AI Governance Framework for Generative AI	30 May 2024	Hallucination, IP, provenance, cyber, training data	Prompt/data policy, provenance, vendor training opt-out, evals for fabrication
Model AI Governance Framework for Agentic AI	22 Jan 2026; update 20 May 2026	Bound the risk, keep humans accountable, technical controls, end-user responsibility	Allowed tools, spend/action limits, kill switch, logs, who is accountable when the agent acts

The agentic framework is the one boards ask about in 2026 because agents plan and act, not only draft. IMDA is explicit: humans stay responsible. Automation bias is named as a failure mode. If your “human in the loop” only clicks approve, you do not have meaningful accountability.

AI Verify: testing, not a licence

AI Verify is IMDA’s testing toolkit and the AI Verify Foundation’s broader programme (including the Global AI Assurance Pilot). It does not replace the PDPA. It gives you process checks and technical tests so “we are responsible” is not an assertion. For a Singapore buyer, an AI Verify report on a material system is more useful than a policy PDF. For an Australian parent, that same pack maps cleanly onto VAISS guardrails 4 (test and monitor) and 9 (records).

PDPA: the part that is already mandatory

Personal data in a prompt, a retrieval index, a fine-tune, a vendor log or a support transcript is still personal data. Purpose limitation, notification, protection and transfer rules do not pause because the processor is a model API. Cross-border transfers to a US or EU model host need a transfer basis. Customer contracts in Singapore now routinely ask where the model runs and whether prompts are used for training.

That is why the inventory fields we use in Australia work here: data in, destination, vendor, training opt-in, owner. See the inventory field list (the hunt is the same; the statute name changes).

MAS and other sector overlays

If you are a bank, insurer or capital-markets entity, MAS technology-risk and outsourcing expectations already cover model APIs as third-party technology. Do not hide an agent behind “innovation”. Put it on the same third-party register as your core processor. Healthcare, education and public-sector buyers will ask for the Model Framework mapping even when the framework is voluntary.

AU + SG groups: one layer, two maps

Evidence field	Australia read	Singapore read
Inventory + owner	VAISS guardrail 1; NAIC register	Model Framework internal governance
Personal data + destination	Privacy Act; Dec 2026 ADM text	PDPA purpose, protection, transfer
Testing	VAISS guardrail 4	AI Verify + agentic pre-deploy tests
Human control	VAISS guardrail 5	Agentic “meaningful accountability”
Records	VAISS guardrail 9; “prove it”	Same artefacts, IMDA language on the cover sheet

Unique insight from dual-market Power Days: the register is not the hard part. The hard part is two legal teams asking for two templates. Give them one system and two export views. Otherwise you will maintain two lies.

Australia context: Can you prove it? · three-regime view: EU vs AU vs SG.

FAQ

Is the Model Framework for Agentic AI mandatory? No. You are still liable for what the agent does under existing law. Counterparties will treat the framework as the expected standard of care.

Do we need AI Verify on every chatbot? No. Use it on systems that face customers, move money, or act with tools. Content-only drafts can stay at policy + inventory.

We only have a Singapore branch and the models sit in Australia. PDPA transfer rules still apply to personal data leaving Singapore. Put residency on the row.

Can a Power Day work in Singapore? Yes. Same day, IMDA vocabulary on the export, PDPA flags instead of APP 1 flags. Detail: Power Day.

Who in the company do we call? The Data Protection Officer. Their contact is public by law. Role and job list: Singapore’s DPO.