PDPA AI Singapore: How PDPC's AI Guidelines Apply to Your Project, With a Checklist
November 10, 2026 · 8 min read · Tim Etzal Group

PDPA AI Singapore compliance starts from one fact: the Personal Data Protection Act 2012 applies to an AI system whenever it collects, uses or discloses personal data, from training and testing through to deployment. PDPC's Advisory Guidelines on use of Personal Data in AI Recommendation and Decision Systems, issued on 1 March 2024, explain how: rely on consent or a recognised exception (such as business improvement or research) to train, notify users clearly and keep written policies when you deploy, and expect good data practice from vendors who build bespoke AI for you.
Updated: 10 November 2026
Is the PDPA mandatory, and does it apply to individuals?
Yes, the PDPA is mandatory for organisations. PDPC describes it as providing "a baseline standard of protection for personal data in Singapore". Personal data is data about an individual who can be identified from that data, or from that data and other information the organisation has or is likely to have access to.
According to PDPC's overview of the Act, the PDPA generally does not apply to:
- an individual acting on a personal or domestic basis;
- an individual acting in their capacity as an employee with an organisation;
- a public agency collecting, using or disclosing personal data;
- business contact information, such as a name, title, business phone number and business email.
So a person using a chatbot for their own errands is outside the PDPA, while a company deploying AI on customer or staff data is inside it.
What PDPC's AI guidelines are, and what they are not
The Advisory Guidelines on use of Personal Data in AI Recommendation and Decision Systems cover systems that embed machine learning models to make decisions or to assist a human decision-maker with recommendations and predictions (paragraph 1.2). They are advisory, "not legally binding" and not legal advice; the PDPA prevails over them (paragraph 2.2). They are organised by project stage:
- Part III, development, testing and monitoring: consent, the business improvement and research exceptions, data protection measures, anonymisation.
- Part IV, deployment: consent, notification and accountability when the AI faces consumers.
- Part V, procurement: good practice for service providers who build bespoke AI with your data.
On 20 July 2026 PDPC added Advisory Guidelines on Use of Personal Data in Generative AI, which build on the 2024 guidelines and are covered further below.
Training and testing AI on personal data (Part III)
Besides consent, an organisation developing AI may rely on two exceptions (paragraph 4.2).
The business improvement exception lets you use personal data you collected in line with the PDPA, without consent, to improve or develop goods, services, methods or processes, to understand customer behaviour and preferences, or to personalise offerings (paragraph 5.1). Two conditions apply: the purpose cannot reasonably be achieved without individually identifiable data, and a reasonable person would consider the use appropriate (paragraph 5.2). PDPC's examples include recommendation engines, job assignment systems for platform workers and internal HR systems that recommend candidates (paragraph 5.4). It can also cover testing an AI system and bias assessment (paragraphs 5.6 and 5.8).
The research exception is for commercial research without a product roadmap. Conditions include a clear public benefit, results that are not used to make decisions affecting the individual, and published results that do not identify anyone (paragraph 6.2).
Whichever basis you use, PDPC recommends data minimisation (7.1), pseudonymisation or de-identification as a basic control where possible (7.2), and a data protection impact assessment when raw personal data must be used (7.3). Anonymised data falls outside the PDPA but still carries re-identification risk (7.8). If you choose personal data over anonymised data, document why, with the decision made at an appropriately senior level (7.9).
Deploying AI to customers (Part IV)
Unless deemed consent or an exception applies, consent is required to collect and use personal data for recommendations, predictions or decisions (paragraph 9.1), and section 20 of the PDPA requires you to notify people of the purpose (9.2). PDPC encourages notifications that explain, to the extent practicable (9.5):
- the product function that needs personal data;
- the types of personal data collected;
- how processing that data is relevant to the function;
- which features of the data are most likely to influence the output.
Notifications can be "layered", with the key points up front and detail in a privacy policy (9.7). If you withhold detail to protect commercially sensitive information, justify and document that decision internally (9.8). If you rely on the legitimate interests exception, for example for AI that detects illegal activity, you must assess that the benefit outweighs any adverse effect and let individuals know you are relying on it (9.9 to 9.11).
Accountability means written policies on how AI uses personal data, in proportion to the risk (10.3), available on request under section 12(d) and ideally published on your website (10.4). Policies can describe fairness measures, data safeguards and, for high-impact outcomes, human oversight (10.6). AI Verify test results can support them (10.10).
Buying AI from a vendor (Part V)
A service provider that builds bespoke AI using your personal data is a data intermediary, with protection and retention obligations (paragraphs 1.4 and 11.2). Good practice includes data mapping and labelling of training data and a provenance record of where the data came from and how it was transformed (11.2). Vendors are encouraged to help you meet your notification, consent and accountability obligations, but PDPC is explicit that organisations "bear the primary responsibility" for ensuring the AI they choose meets those obligations (11.8). Part V does not cover off-the-shelf products (11.1); the 2026 guidelines do.
What the 2026 generative AI guidelines added
- AI-specific notice for training. General statements such as "new product development" are not a sufficient means of obtaining consent for large-scale generative AI training or fine-tuning; you must say AI training is a purpose (paragraph 4.4), ideally with an easy way to decline (4.6).
- No forced consent. You must not require consent beyond what is reasonable to provide the service (4.8).
- Clear roles. PDPC names model providers, system providers and system deployers (6.2). A company buying a generative AI tool is a system deployer: it bears primary responsibility for PDPA compliance of the tool it chooses, and needs enough information on upstream safeguards when procuring (9.1).
- New data to protect. Prompts, generated outputs and agent or tool activity data must be safeguarded, with owners assigned (9.3), and safeguards reviewed as agentic features appear (9.5).
- Access and correction. Requests still apply to data used in generative AI, for example data held in a retrieval database (10.3 and 10.5).
PDPA AI Singapore checklist by project stage
| Stage | Check | Source |
|---|---|---|
| Before data is used | Identify the personal data, minimise it, pseudonymise where possible | AI guidelines 7.1 to 7.2 |
| Before data is used | Record the basis: consent, business improvement or research exception | AI guidelines 4.2, 5.1, 6.2 |
| Before data is used | Run a DPIA if raw personal data is needed | AI guidelines 7.3 |
| Procurement | Ask the vendor for data mapping, provenance records and safeguards | AI guidelines 11.2, GenAI guidelines 9.1 |
| Procurement | Confirm what the vendor retains and for how long | GenAI guidelines 7.2 to 7.3 |
| Deployment | Notify users of function, data, relevance and key features | AI guidelines 9.5 |
| Deployment | Use AI-specific notice and an opt-out for generative AI training | GenAI guidelines 4.4, 4.6 |
| Deployment | Publish written policies proportionate to the risk | AI guidelines 10.3 to 10.4 |
| Operation | Assign owners for prompts, outputs and agent activity data | GenAI guidelines 9.3 |
| Operation | Handle access and correction requests that touch AI data | GenAI guidelines 10.3 |
PDPA records fit into the wider governance work described in our guide to the AI governance framework Singapore publishes, which covers the voluntary IMDA frameworks around them.
Key takeaways
- The PDPA is mandatory for organisations and applies to AI whenever personal data is involved; individuals acting in a personal or domestic capacity are outside it.
- PDPC's AI guidelines (1 March 2024) are advisory, but they show how PDPC reads the binding Act.
- To train on personal data, rely on consent or the business improvement or research exception, and document which.
- At deployment, notify users in plain language and keep written policies proportionate to the risk.
- Vendors are data intermediaries, but the organisation using the AI bears primary responsibility.
PDPA and AI projects with Etzal Group
Etzal Group's AI Consulting package maps each AI use case against these sections and hands over the checklist above, filled in, with the use-case inventory and risk tiering. We are not lawyers: the legal answer for your case stays with your data protection officer or counsel. If you then build with a vendor, us included under Enterprise Software, ask for the data mapping and provenance records Part V describes. Still choosing? Our build vs buy AI agents guide and what an AI readiness assessment should deliver help, or book a consultation.
Frequently asked questions
Is the PDPA mandatory?
Yes, for organisations. The Personal Data Protection Act 2012 is law and sets a baseline standard for how organisations in Singapore collect, use, disclose and care for personal data, including inside AI systems.
Does the PDPA apply to individuals?
Generally not when an individual acts on a personal or domestic basis, or as an employee acting for an organisation. In the employee case, the organisation itself carries the obligations.
Can a company use customer data to train AI without consent in Singapore?
Sometimes. PDPC's 2024 guidelines explain that the business improvement exception can apply if the purpose cannot reasonably be achieved without identifiable data and a reasonable person would consider the use appropriate. If you rely on consent for large-scale generative AI training, the 2026 guidelines require an AI-specific notification rather than a general one.
Are PDPC's AI guidelines legally binding?
No. Both the 2024 and 2026 guidelines say they are advisory and not legally binding, but they explain how PDPC applies the PDPA, which is binding.
Who is responsible if a vendor's AI mishandles personal data?
The vendor has protection and retention obligations as a data intermediary, but PDPC states that the organisation choosing the AI bears primary responsibility for meeting its PDPA obligations.
Do we have to tell customers that we use AI?
Where consent is needed, the PDPA requires you to notify people of the purpose. PDPC encourages explaining the function, the data used, why it is relevant and which data most influences the output, in plain language.
Related service
WhatsApp AI Chatbot
We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.
