Part of the AI Regulation News hub.
Thailand's Personal Data Protection Committee issues version 1.0 of its non-binding AI guidelines, treating prompts, logs and embeddings as potential personal data
The guidelines say they bind nobody. Their value is that the regulator sets out how it reads the PDPA for chatbots, RAG systems and vendor AI, in a document organisations can be measured against.
Bottom line: This is regulator guidance, not a statute. The content edition says on its face that it has no legal enforcement force, does not bind decisions on offences, and is not a definitive interpretation of the PDPA. It was issued by the Personal Data Protection Committee under section 16(3) of the PDPA, according to the cover, and the PDPC page dates it 7 October 2026.
Who this affects: Thai data controllers and processors that build or deploy AI, including chatbots, retrieval-augmented systems and AI agents, along with data protection officers, in-house counsel and procurement teams that sign AI vendor contracts.
Issue date: PDPC page dated 7 October B.E. 2569 (7 October 2026). The first-page header of the full announcement PDF reads Royal Gazette, 2 November B.E. 2569 (2 November 2026). We could not read the announcement's own signing date, so we give both and rely on neither as the issue date.
What changed: Version 1.0 of the guidelines, as an announcement and a content edition, now sits on the PDPC site; the content edition's own running headers still read "Draft of Guidelines", a discrepancy we could not resolve. On its executive summary, prompts, logs, embeddings and outputs may be personal data, a trained model is not automatically anonymous, roles turn on decision power, consent is a last resort, high-risk projects should have a DPIA, and technical difficulty does not excuse erasure duties.
Analysis: On our reading, the vendor-role example matters most for contracts. The guideline says a cloud AI provider acting as processor becomes a further controller for the part of its activity where it trains its own model on customer prompts. That points procurement teams at the terms-of-service clause on training rather than at the label on the contract. The contract advice is our inference from the guideline's example.
Primary sources: PDPC page for the announcement and its two downloads (Thai) · Content edition PDF, 75 pages (Thai, as extracted)
- Instrument (EN)
- PDPC Announcement: Guidelines on personal data protection in the development and use of AI, B.E. 2569 (2026), version 1.0 (our rendering of the Thai title)
- Instrument (native)
- Thai-language announcement and guidelines on the PDPC site; see the source link
- Authority
- Personal Data Protection Committee (PDPC), prepared with the PDPC Office and the Electronic Transactions Development Agency (ETDA), per the content edition cover
- Jurisdiction
- Thailand
- Status
- Version 1.0 posted by the PDPC; in the content edition, running headers still read "Draft of Guidelines", an unreconciled discrepancy
- Bindingness
- Guidance. The content edition says it has no legal enforcement force and does not bind decisions on offences
- Issue date / next deadline
- PDPC page dated 7 October 2026; Royal Gazette header on the announcement PDF reads 2 November 2026; no deadline found
- Primary source
- https://www.pdpc.or.th/28888/
What is the document, and what does it say about itself?
The PDPC page lists two downloads: the announcement, and the content edition of the guidelines. Our extraction of the announcement PDF shows 81 pages, a cover reading version 1.0, and a first-page header that reads as the Royal Gazette of 2 November B.E. 2569. The text layer of the gazette pages came out garbled, so we could not read the announcement's operative wording, its signing date or any effective-date clause.
The content edition is legible. Its preface says it is initial advice for the benefit of controllers and others, and that it is ไม่มีสภาพบังคับตามกฎหมายและไม่มีผลผูกพันในการพิจารณาความผิด (in our rendering, without legal enforcement force and not binding in deciding offences). It covers analytic, generative and agentic AI across the AI lifecycle.
When does an AI input or output count as personal data?
The executive summary says prompts, logs, embeddings and generated outputs may all be personal data under section 6 if a person can still be identified directly or indirectly. It follows that a model trained on personal data is not automatically anonymous: โมเดลที่ฝึกด้วยข้อมูลส่วนบุคคลจึงไม่ถือเป็นข้อมูลนิรนามโดยอัตโนมัติ (a model trained on personal data is therefore not automatically treated as anonymous data, our rendering), unless the test in section 2.4 of the guidelines is met.
We read the summary and the table of contents. We did not test every sub-section against the summary, and the anonymisation criteria in section 2.4 are not described here.
Who is controller and who is processor?
The guidelines tell an organisation to identify, for each activity in the AI value chain, whether it is controller or processor, by looking at who decides the purposes and means rather than at a technical job title. The executive summary flags the case where a provider takes customer data to train its own model, which makes it a separate controller for that use.
The worked table is concrete. A deployer using an off-the-shelf tool to rank job applicants decides the purpose and so acts as controller. A developer that scrapes public web pages for a training set is its own decision-maker. A cloud AI provider that is a processor for the service, but whose terms let it use customer prompts to fine-tune its own model, becomes a further controller for that activity.
What does it say on legal basis, DPIAs and erasure?
On legal basis, the summary says consent should be the last resort, not the first, and that an organisation should work through the section 24 or section 26 bases first. Legitimate interest needs a written legitimate interest assessment. Relying on consent alone creates operational risk when consent is withdrawn.
On risk, projects likely to affect individuals, such as automated decisions with an effect on rights or training on sensitive data, should go through a data protection impact assessment before they start, alongside AI-specific safeguards such as input and output guardrails, limits on agent permissions, and logging that supports later forensic work. The guideline words these as expectations in the plain sense of "should".
On erasure, the summary says technical limits such as machine unlearning do not remove the legal duty. For data held in a vector database or RAG index, the guideline says the index entry can be deleted directly. For data embedded in a foundation model, where deletion within 90 days is not feasible for technical or cost reasons, it suggests equivalent measures such as output filtering and guardrails until the next model update removes the data permanently, in line with the PDPC erasure rules.
The summary also names two timed duties under the PDPA itself: notice of a breach to the PDPC within 72 hours of becoming aware under section 37(4), and notice to individuals where the risk is high.
Why do the two dates matter?
The PDPC page is dated 7 October B.E. 2569, which is 7 October 2026, and its publication timestamp is 7 October 2026 at 16:00 Thailand time. The same page shows a last-modified time of 3 November 2026, so it was edited after posting. Separately, the announcement PDF's first-page header, as we extracted it, reads Royal Gazette, 2 November B.E. 2569.
Our inference, not the PDPC's statement, is that the announcement was posted on the website on 7 October and appeared in the Gazette on 2 November. An organisation looking for the date from which the document counts as official should read the Gazette page itself in Thai.
What we did not verify
What we opened: the PDPC announcement page, the 75-page content edition PDF (read as extracted Thai text, including the preface, executive summary, the role table and the erasure section), and the first pages of the 81-page announcement PDF. Thai quotations are reproduced in the original characters, and every English rendering is ours.
What we did not open or could not read: the operative text of the announcement, its signing date and any effective-date clause, because the gazette pages extracted as garbled text. We read the 7 October date from the website page, and the 2 November date from the header text, which we transcribed by hand and did not match against a saved file. We did not read the full 75 pages line by line. We did not open the earlier draft, so we make no claim about its date or what changed from it.
Also unresolved: the content edition PDF carries "Draft of Guidelines" in its running headers and cover subtitle while the PDPC page lists it beside the final announcement. We could not confirm that the content edition matches the final announcement text, so quotations from it describe that PDF and not necessarily the gazetted version. We also did not verify that the content edition is the PDF named "content edition" on the page.
What we refuse to claim: that any organisation is in breach for departing from the guidelines, that the PDPC will enforce them, or what section numbers beyond those quoted say. Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
A Thai controller that uses outside AI should read its vendor terms for one thing first: whether customer prompts can be used to train the vendor's own model. The guideline treats that as a change of role. After that, the practical file is short: a map of where prompts and logs are stored, a DPIA for any high-risk use, and a deletion route that works for the vector index and has an interim answer for the model itself.
Source File
Open the PDPC page and note the date 7 October B.E. 2569 and the two downloads. Open the announcement PDF and read the first-page header and cover. Open the content edition and read the preface, then the executive summary, then the table of contents and the erasure section.
ไม่มีสภาพบังคับตามกฎหมายและไม่มีผลผูกพันในการพิจารณาความผิด (our rendering: without legal enforcement force and not binding in deciding offences) · PDPC AI guidelines, content edition, preface
FAQ
Are Thailand's PDPC AI guidelines legally binding?
The content edition says they have no legal enforcement force and do not bind decisions on offences. The PDPA itself still applies, including the 72-hour breach notice the guidelines point to.
Can prompts and logs be personal data?
The executive summary says prompts, logs, embeddings and generated outputs may be personal data if the person can still be identified directly or indirectly.
Does an AI vendor become a controller if it trains on customer data?
On the guideline's worked example, a provider that is a processor for the service becomes a separate controller for the activity where it uses customer data to train its own model.
Is consent the main legal basis for AI training?
The summary says consent should be the last resort and that organisations should first analyse the section 24 and section 26 bases, with a legitimate interest assessment where that basis is used.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.