Part of the AI Regulation News hub.
Singapore's PDPC closed its consultation and issued the final Advisory Guidelines on Use of Personal Data in Generative AI, dropping the proposed web-scraping notification practice.
The draft told organisations to tell a website owner before scraping it. Respondents pushed back on both workability and effect, and the Commission removed the practice, replacing it with a documentation duty that is harder to ignore.
Bottom line: Advisory Guidelines. They are guidance on how the Commission reads the PDPA, not a new statute and not an amendment to one. The consultation is now closed and the final Guidelines are published.
Who this affects: Data protection officers, privacy counsel and AI product owners at organisations that train, fine-tune, host or deploy generative models involving personal data of individuals in Singapore.
Issue date: 20 July 2026 for the closing note and the final Guidelines. The consultation opened 2 June 2026 and closed 1 July 2026. The response document is dated Issued 20 July 2026.
What changed: The recommended practice of notifying a controlling organisation before web scraping was removed. A written-record duty was added where reliance on the Publicly Available Exception is arguable. AI-Specific Notifications were clarified as a sufficiency standard, not a template.
Analysis: The removed practice was replaced with something less comfortable. Dropping the notice obligation looks like relief until you read what took its place: if it is reasonably arguable that data behind a digital barrier is not publicly available and you scrape it anyway, you must record your reasoning and produce it if the Commission asks.
Primary sources: Consultation page with the 20 July 2026 closing entry · Closing note / response to feedback (PDF) · Final Advisory Guidelines landing page
- Instrument (EN)
- Advisory Guidelines on Use of Personal Data in Generative AI
- Authority
- Personal Data Protection Commission (PDPC), supported by IMDA
- Jurisdiction
- Singapore
- Status
- Final. Consultation closed 1 July 2026, response and Guidelines issued 20 July 2026.
- Bindingness
- Advisory. Guidance on the Commission's reading of the Personal Data Protection Act 2012; it does not amend the Act.
- Issue date / next deadline
- 20 July 2026. No further deadline set.
- Consultation responses
- 40 organisations and three individuals
- Read alongside
- Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems; Advisory Guidelines on Key Concepts in the PDPA
- Primary source
- https://www.pdpc.gov.sg/organisations/regulations-decisions/regulatory-guidance/public-consultation-on-the-proposed-advisory-guidelines-on-use-of-personal-data-in-generative-ai
What was consulted on and what closed
On 2 June 2026 the Commission opened a consultation on proposed Advisory Guidelines covering three questions: how personal data may be collected and used to develop generative models, how protection responsibilities are allocated across the generative AI lifecycle, and how individuals' requests about their data should be handled.
The consultation closed on 1 July 2026 with 40 responses from organisations across technology, finance, healthcare and entertainment, plus three individuals. On 20 July the Commission issued its response to feedback and published the final Guidelines in alignment with it.
Our earlier coverage of this instrument described the 2 June draft. This page now reflects the final text and the Commission's account of what it changed.
The scraping notice that did not survive
The draft carried a recommended best practice: an organisation intending to web-scrape should notify the organisation controlling the site. Respondents said this would slow development and was hard to operate. Who in the controlling organisation should be told? How long should you wait before proceeding?
There was a sharper objection too. Respondents warned that a controlling organisation, once notified, could use the warning to impose fresh access restrictions, which would defeat reliance on the Publicly Available Exception in the first place.
The Commission removed the practice. In its place it expanded the list of examples of digital barriers, added illustrative box stories, and set out two clarifications.
What replaced it
First, on the Commission's account, where it is reasonably arguable that data behind a digital barrier is not publicly available but an organisation relies on the exception regardless, it must explain its assessment and reasoning in a Data Protection Impact Assessment or other written records, and produce that documentation if the Commission requires it.
Second, the Commission pushed a duty back onto publishers: organisations and individuals who put personal data online should implement appropriate digital barriers where they do not intend, or have not obtained consent, for that data to be scraped.
Respondents had asked whether token-based access, API rate limits, age gates, geo-blocking and robots.txt count as digital barriers. The closing note does not answer that question item by item. What it records is that the Commission expanded its list of examples of digital barriers and added illustrative box stories.
Notifications, roles and agents
Respondents supported the intent of AI-Specific Notifications but split on their scope and form. The Commission clarified that the phrase is shorthand for giving enough information about the use of User Data to enable meaningful consent, and that no list-style template or standard format is required. Existing privacy policies and terms of use can carry the load, provided individuals can understand what data is used and how.
It also narrowed a definition that had been read too widely. Training and fine-tuning refer only to activities that develop or modify a model's parameters or underlying capabilities. Inference and operational activities, and Retrieval-Augmented Generation specifically, fall outside.
On roles, the Commission accepted that a single entity may be Model Provider, System Provider and System Deployer at once, and said such stakeholders must hold policies and practices sufficient to meet each set of obligations. On agentic systems, respondents raised persistent memory, disclosure through external connectors, and unclear allocation across multi-agent chains. The Commission cited additional agent capabilities and risks but said a comprehensive treatment sits outside these Guidelines and will be the subject of separate study.
The access and correction problem nobody solved
Respondents made a practical objection, and the Commission responded by refining its list of best practices: it is impractical to expect Model Providers processing heterogeneous datasets at scale to document a comprehensive lineage of training data. A few responses also warned that machine unlearning is still an emerging field and premature to cite as a measure organisations should track and adopt.
The Commission refined its best practice list and clarified scope: the Access and Correction Obligations apply to personal data collected, used or disclosed for model or system development or deployment. It also set out considerations for different stakeholders responding to individual requests, after respondents pointed out that a System Provider typically has neither possession of nor control over model training data.
That leaves a familiar gap. An individual's correction request can be valid and still be answerable only through output safeguards rather than by changing what a model learned.
What we did not verify
We opened the consultation page, the final Advisory Guidelines landing page, and the full seven-page closing note titled Response to Feedback on Public Consultation on Proposed Advisory Guidelines on Use of Personal Data in Generative AI, issued 20 July 2026.
We did not open the final Guidelines document itself, the 2 June proposed Guidelines, the cover note, or the published responses from the 43 respondents. Our account of what changed is therefore the Commission's own summary of its changes, not a diff we performed against the two texts.
We will not claim any specific paragraph number in the final Guidelines, nor that the wording of the removed scraping practice is absent in every form. The closing note says the best practice was removed; we have not read the final text to confirm the drafting.
Advisory Guidelines carry no penalty of their own, but they tell you what the regulator will ask for when it does look. The concrete change here is evidentiary: if your scraping relies on the Publicly Available Exception in a case where the barrier question is genuinely arguable, the Commission expects a written assessment on file. Write that assessment now, at collection time, because it is not reconstructable afterwards.
Source File
Open the PDPC consultation page and confirm the 20 July 2026 entry stating that a closing note was issued and the Advisory Guidelines published in alignment. Then open the closing note PDF and confirm paragraph 1.3 records 40 organisation responses and three individuals, and paragraph 2.3 states the notification best practice was removed.
Having considered respondents' feedback, the Commission has removed the best practice of notifying controlling organisations from the Guidelines. ยท PDPC, Response to Feedback, 20 July 2026
FAQ
Are these Guidelines binding?
No. They are Advisory Guidelines setting out how the Commission reads the Personal Data Protection Act 2012. They do not amend the Act and impose no separate penalty.
Do I still have to notify a website before scraping it?
The Commission removed that recommended practice after consultation feedback. What replaced it is a record-keeping expectation where reliance on the Publicly Available Exception is reasonably arguable.
Does an AI-Specific Notification need a set format?
The Commission clarified that it does not. The test is whether individuals can understand what User Data is used and how, and existing privacy policies or terms of use may carry that information.
Is Retrieval-Augmented Generation treated as training?
No. The Commission clarified that training and fine-tuning cover only activities that develop or modify a model's parameters or underlying capabilities, and that inference and operational activities such as RAG are not included.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.