The guidance is advisory rather than a standalone ban on AI training. It works within Singapore’s Personal Data Protection Act (PDPA), including exceptions that may allow some uses without consent. One important example is the treatment of certain publicly available personal data used for AI development.

What an AI-specific notification should explain

A useful notice should give people enough information to understand both the product and the data pipeline. The proposed approach called for organisations to explain:

• What the model does and why it is being developed
• Which categories of personal data are used, such as contact details, transaction or location information, video, or voice recordings
• How the data contributes to the model, rather than merely saying that it is processed
• How a person can opt out or withdraw consent, using practical and reasonably accessible instructions

The PDPC’s concern is transparency that works at the point of decision—not a buried sentence in a lengthy privacy policy. Suggested formats included a prominent in-product notice or a dedicated webpage.

Example: voice data for text-to-speech

A company developing text-to-speech functionality could explain that it uses customers’ voice recordings to train the model and that the recordings help the system recognise speech patterns. That is materially clearer than describing the purpose only as product improvement.

Notification is not the same as consent

The most important compliance distinction is between telling people what will happen and establishing a lawful basis for using their data. Under the PDPA’s general framework, organisations generally need to notify individuals of the purpose and obtain consent unless an exception applies.

The final guidance clarifies the content of the notification when consent is required: the organisation must expressly address the intended AI model-development use. That does not mean every use of personal data for AI training automatically requires affirmative, explicit consent. The answer can depend on the applicable PDPA exception, including rules concerning publicly available information.

For businesses, the practical lesson is to document the legal basis for each training dataset separately. A clear notice cannot replace consent where consent is required, and consent cannot be assumed from language too general to describe the actual AI use.

Questions raised by the draft proposal

The consultation-stage proposal attracted attention because several operational questions were not obvious from the initial notice requirement.

Is explicit consent always required?

The draft did not clearly establish that affirmative, explicit consent would be required in every GenAI-training scenario. It instead prompted questions about when an opt-out approach might be acceptable and when an existing PDPA exception could apply.

The final guidance’s treatment of consent and exceptions makes the legal basis more important, not less. Organisations should avoid presenting an AI-specific notice as a universal substitute for consent.

Does anonymisation remove the AI-specific notice issue?

The draft also left uncertainty about whether the special notification expectation would apply when data were anonymised before training. That question matters because the privacy analysis may differ depending on whether information can still be linked to an identifiable individual and how the organisation performs and documents the anonymisation.

The available reporting does not support treating anonymisation as an automatic answer for every dataset. Companies should assess the data and processing method rather than assuming that labelling a dataset “anonymised” ends the analysis.

Can a business deny a service to someone who refuses?

Another unresolved issue was whether banks, insurers, platforms, or other providers could withhold or make a service unavailable when an individual refused the use of their data for AI training. This is where formal choice can meet practical pressure: an opt-out may exist on paper but be difficult to exercise if refusing it carries a significant cost.

The notification requirement improves visibility, but it does not by itself answer every question about fairness, service access, or the bargaining power between individuals and providers.

Why AI-enabled devices are part of the wider privacy debate

The PDPC’s attention to generative AI sits within a broader concern about products that collect data in ways people may not notice. Reporting around Commissioner Denise Wong’s priorities has highlighted devices and systems such as smart glasses, smartwatches, and palm-scanning payment systems.

These products can involve sensitive information, including facial features, fingerprints, voice data, and palm-vein patterns. The reported risks include covert recording, exposure or misuse of biometric information, and smart glasses being used to capture examination material or other information to facilitate cheating.

Singapore’s Ministry of Digital Development and Information has also addressed questions about visual indicators on smart glasses and the collection of biometric or environmental data from people nearby who are not using the device.

The underlying issue is meaningful awareness. A small recording light or a device policy hidden in documentation may not give bystanders a realistic opportunity to understand what is happening or make a choice.

What organisations should do now

Teams developing or deploying GenAI systems in Singapore should treat the guidance as a data-governance checklist:

1. Map each dataset to a specific AI use case. Identify whether the data are used for training, fine-tuning, testing, monitoring, or another purpose.
2. Choose and document the PDPA basis. Record whether consent is required or whether a statutory exception is being relied on.
3. Rewrite generic notices. State plainly that personal data will be used to develop or improve a generative-AI model where that is the purpose.
4. Describe data categories and model function. People should be able to understand what information is used and what it helps the model do.
5. Make choice usable. Explain how to opt out or withdraw consent, and record how requests are handled.
6. Review connected devices separately. Smart glasses, wearables, cameras, and biometric systems may create notice and consent questions involving people who are not the direct user.

Singapore’s shift is therefore less about banning the use of personal data in generative AI than about making the use visible, specific, and tied to an identifiable legal basis. The final guidance closes some of the gap between traditional privacy notices and modern AI development, while the hardest questions—especially meaningful choice and the treatment of sensitive or indirectly captured data—remain active governance concerns.

The guidance is advisory rather than a standalone ban on AI training. It works within Singapore’s Personal Data Protection Act (PDPA), including exceptions that may allow some uses without consent. One important example is the treatment of certain publicly available personal data used for AI development.

What an AI-specific notification should explain

A useful notice should give people enough information to understand both the product and the data pipeline. The proposed approach called for organisations to explain:

• What the model does and why it is being developed
• Which categories of personal data are used, such as contact details, transaction or location information, video, or voice recordings
• How the data contributes to the model, rather than merely saying that it is processed
• How a person can opt out or withdraw consent, using practical and reasonably accessible instructions

The PDPC’s concern is transparency that works at the point of decision—not a buried sentence in a lengthy privacy policy. Suggested formats included a prominent in-product notice or a dedicated webpage.

Example: voice data for text-to-speech

A company developing text-to-speech functionality could explain that it uses customers’ voice recordings to train the model and that the recordings help the system recognise speech patterns. That is materially clearer than describing the purpose only as product improvement.

Notification is not the same as consent

The most important compliance distinction is between telling people what will happen and establishing a lawful basis for using their data. Under the PDPA’s general framework, organisations generally need to notify individuals of the purpose and obtain consent unless an exception applies.

The final guidance clarifies the content of the notification when consent is required: the organisation must expressly address the intended AI model-development use. That does not mean every use of personal data for AI training automatically requires affirmative, explicit consent. The answer can depend on the applicable PDPA exception, including rules concerning publicly available information.

For businesses, the practical lesson is to document the legal basis for each training dataset separately. A clear notice cannot replace consent where consent is required, and consent cannot be assumed from language too general to describe the actual AI use.

Questions raised by the draft proposal

The consultation-stage proposal attracted attention because several operational questions were not obvious from the initial notice requirement.

Is explicit consent always required?

The draft did not clearly establish that affirmative, explicit consent would be required in every GenAI-training scenario. It instead prompted questions about when an opt-out approach might be acceptable and when an existing PDPA exception could apply.

The final guidance’s treatment of consent and exceptions makes the legal basis more important, not less. Organisations should avoid presenting an AI-specific notice as a universal substitute for consent.

Does anonymisation remove the AI-specific notice issue?

The draft also left uncertainty about whether the special notification expectation would apply when data were anonymised before training. That question matters because the privacy analysis may differ depending on whether information can still be linked to an identifiable individual and how the organisation performs and documents the anonymisation.

The available reporting does not support treating anonymisation as an automatic answer for every dataset. Companies should assess the data and processing method rather than assuming that labelling a dataset “anonymised” ends the analysis.

Can a business deny a service to someone who refuses?

Another unresolved issue was whether banks, insurers, platforms, or other providers could withhold or make a service unavailable when an individual refused the use of their data for AI training. This is where formal choice can meet practical pressure: an opt-out may exist on paper but be difficult to exercise if refusing it carries a significant cost.

The notification requirement improves visibility, but it does not by itself answer every question about fairness, service access, or the bargaining power between individuals and providers.

Why AI-enabled devices are part of the wider privacy debate

The PDPC’s attention to generative AI sits within a broader concern about products that collect data in ways people may not notice. Reporting around Commissioner Denise Wong’s priorities has highlighted devices and systems such as smart glasses, smartwatches, and palm-scanning payment systems.

These products can involve sensitive information, including facial features, fingerprints, voice data, and palm-vein patterns. The reported risks include covert recording, exposure or misuse of biometric information, and smart glasses being used to capture examination material or other information to facilitate cheating.

Singapore’s Ministry of Digital Development and Information has also addressed questions about visual indicators on smart glasses and the collection of biometric or environmental data from people nearby who are not using the device.

The underlying issue is meaningful awareness. A small recording light or a device policy hidden in documentation may not give bystanders a realistic opportunity to understand what is happening or make a choice.

What organisations should do now

Teams developing or deploying GenAI systems in Singapore should treat the guidance as a data-governance checklist:

1. Map each dataset to a specific AI use case. Identify whether the data are used for training, fine-tuning, testing, monitoring, or another purpose.
2. Choose and document the PDPA basis. Record whether consent is required or whether a statutory exception is being relied on.
3. Rewrite generic notices. State plainly that personal data will be used to develop or improve a generative-AI model where that is the purpose.
4. Describe data categories and model function. People should be able to understand what information is used and what it helps the model do.
5. Make choice usable. Explain how to opt out or withdraw consent, and record how requests are handled.
6. Review connected devices separately. Smart glasses, wearables, cameras, and biometric systems may create notice and consent questions involving people who are not the direct user.

Singapore’s shift is therefore less about banning the use of personal data in generative AI than about making the use visible, specific, and tied to an identifiable legal basis. The final guidance closes some of the gap between traditional privacy notices and modern AI development, while the hardest questions—especially meaningful choice and the treatment of sensitive or indirectly captured data—remain active governance concerns.