TL;DR
- Thailand's PDPC office announced on 16 November 2026 that the Royal Gazette published the Committee's AI personal data guidelines, version 1.0, in an 81-page gazette file.
- The notification is dated 22 October 2026 and was gazetted on 2 November 2026; the PDPC says it is preliminary advice with no legal force and is not a ruling on the PDPA.
- Section 4.1 says prompts, user context, telemetry logs and embeddings can still be personal data if they can be linked back to a person.
- Section 4.5 says consent should be the last option and that legitimate interest under PDPA section 24(5) needs an assessment with purpose, necessity and balancing tests.
- The executive summary stresses reporting breaches to the PDPC within 72 hours under section 37(4) and running a DPIA before high-risk AI projects start.
Thailand's Office of the Personal Data Protection Committee announced on 16 November 2026 that the Royal Gazette has published the Committee's guidelines on personal data protection in the development and use of artificial intelligence, according to the PDPC announcement. The notification is dated 22 October 2026, appears in Royal Gazette volume 143, special part 261 Ngor of 2 November 2026, and attaches a guideline labelled version 1.0. The PDPC states that the document is preliminary advice with no legal force. It still sets out in writing the practice the Committee recommends under the Personal Data Protection Act B.E. 2562 (PDPA) for prompts, chatbots, model training and AI vendors.
This article summarises the parts of the guideline that were read for it and cites section numbers throughout. It is not legal advice.
What the PDPC published and its legal status
The gazette notice is titled, in translation, "Notification of the Personal Data Protection Committee on guidelines for personal data protection in the development and use of artificial intelligence technology B.E. 2569". It cites section 16(3) of the PDPA and resolutions of Committee meetings 8/2569 on 26 August 2026 and 9/2569 on 24 September 2026, and is signed by Committee chairman Thienchai Na Nakorn. The cover of the attached guideline says it was issued by the Committee and prepared by the PDPC office together with the Electronic Transactions Development Agency (ETDA). The gazette PDF runs to 81 pages.
On legal status, the PDPC announcement and the guideline's own preamble use the same wording. Translated closely, the guideline "is intended as preliminary advice, solely for the benefit of data controllers or related persons in applying it. It has no legal force, is not binding in the consideration of an offence, and is not to be regarded as an interpretation or ruling on problems arising from enforcement of the law on personal data protection."
The executive summary adds that these are technology-specific guidelines to be used alongside the PDPA, its subordinate rules and the PDPC office's other basic guidelines, and that they do not replace any of them. The pages read for this article give no penalties of their own and no separate effective date, and do not say whether a binding notification will follow.
Who and what the guideline covers
The executive summary says the guideline covers Analytic AI, Generative AI and Agentic AI across the AI lifecycle, from planning and data preparation through model training and deployment to retirement. Section 2.2 says it applies to collection, use or disclosure of personal data by data controllers and data processors under the PDPA, and not to personal or household use under section 4(1).
Section 2.2 also refers to the territorial scope in PDPA section 5, which section 2.3 restates. A controller or processor located outside Thailand that collects, uses or discloses data of people in Thailand must comply, in particular when offering goods or services to people in Thailand or monitoring their behaviour in the country. Section 2.3 also draws the line between personal and organisational use: an individual who uses other people's personal data with a generative AI system for commercial gain cannot rely on the personal-use exemption.
Five positions that matter in day-to-day AI use
The table below lists five positions taken in sections of the guideline that were read for this article.
| Topic | What the guideline says | Section |
|---|---|---|
| What counts as personal data | Prompts, user context, telemetry logs and vector embeddings can remain personal data if they can be processed or matched back to identify a person. | 4.1 |
| Trained models | A model trained on personal data is not automatically anonymous. The controller should assess the chance of extracting training data from the model and of obtaining it through prompts. | 2.4 |
| Controller or processor | Status depends on who decides the purposes and means. An AI provider that uses the deploying organisation's personal data for its own benefit, such as training or fine-tuning, becomes a controller for that activity. | 3.1 |
| Lawful basis | Consent is not the only basis and should be the last choice. Legitimate interest under section 24(5) needs a documented assessment with purpose, necessity and balancing tests. | 4.5 |
| Sending data to an AI provider | Entering or uploading personal data into an external AI service may be a disclosure to the provider, and a cross-border transfer if people abroad can access it. | 5.3 |
Prompts, logs and embeddings
Section 4.1 tells controllers to look at the technical data an AI process creates. It names prompts, user context and telemetry logs, and adds vectors and embeddings as items that are often overlooked. Humans cannot read an embedding directly, the section says, but if the data can be compared or processed back to identify a person it may still be personal data protected by law. The guideline's own FAQ repeats the point for embeddings (item 1.3) and for AI outputs such as a realistic deepfake of a living person (item 1.2).
Controller and processor roles with AI vendors
Section 3.1 says an organisation that deploys an AI system for its own business purposes is generally a data controller under PDPA section 6, because it decides what personal data goes in and what the output is used for. A developer or AI provider is a processor only where it acts on the controller's instructions without power over purposes and means. A warning box in the same section covers the common mixed case: a model provider that also uses the deploying organisation's personal data for its own benefit, for example training or fine-tuning to serve other customers, becomes a separate, independent controller, and the contract should set the scope of authority clearly. Appendix B, the checklist for organisations using AI, asks for a data processing agreement under section 40 that bars the provider from using the organisation's personal data beyond the agreement unless permitted.
Lawful basis, legitimate interest and consent
Section 4.5 says organisations should recognise that consent is not the only legal basis and should be the last option when no other basis applies. Its reasoning is operational: if a data subject withdraws consent, the organisation may have to stop processing or even remove the data from a training set, which can affect the model. For general personal data, the section recommends considering legitimate interest under section 24(5), on condition that the organisation carries out a Legitimate Interest Assessment (LIA) built on three tests: purpose, necessity and balancing. It also warns that a controller that chose consent for an activity cannot fall back on section 24(5) for the same activity after the data subject refuses or withdraws.
Where consent is used, section 4.6 sets out how. Consent must be freely given, and the section uses a chatbot as its example: an organisation should not make consent to using conversations for model training a condition of using the chatbot, and should offer that choice separately. Consent must be specific, so broad purposes such as "to improve the user experience" should be avoided, and boxes should not be pre-ticked. Withdrawal must be as easy as giving consent.
Reusing customer data and sending it abroad
Section 5.1 applies the purpose limitation rule in PDPA section 21 to AI. Its worked example is customer service data, such as call recordings or chat messages collected to resolve problems and improve service, being reused to train a new product such as a generative AI chatbot. The organisation should consider carefully whether that is a different purpose from the one originally notified, and if so notify the new purpose or anonymise the data first.
Section 5.3 deals with third-party AI providers, including cloud services, models offered through an API and software as a service. Entering prompts or uploading personal data so the provider's system can process them may be regarded as disclosure to that provider, which may store the data in data centres in Thailand or abroad. The section says mere transit or storage with no access by outside persons is not treated as a cross-border transfer. If the provider stores or uses the data in a data centre abroad and persons there other than the sender can access it, that may be a transfer, and PDPA sections 28 and 29 and the related notifications apply.
Automated decisions, objections and deletion
Section 5.4 covers AI outputs that support or replace decisions with legal or similarly significant effects on a person, with credit approval, credit scoring, recruitment and state benefits as examples. Controllers should tell data subjects in advance, explain the logic concisely, and cannot use trade secrets to refuse any explanation at all. Human review counts as meaningful only if the reviewer considers the output with other information, has real authority and competence to overturn it, and is trained against automation bias. Approving whatever the system says is described as rubber-stamping.
Section 9.3 notes that the PDPA right to object under section 32 applies especially where processing is for direct marketing or rests on legitimate interest, which in AI often takes the form of profiling. Controllers should be able to separate out and stop using a person's data when the right is exercised. Section 9.2 addresses erasure under section 33. It recommends keeping data separate from the model, for example in a retrieval-augmented generation (RAG) setup where an individual's entries can be deleted from the vector database. Where removing data from a trained model is not technically feasible, the section points to output filtering or guardrails until the next retraining cycle.
The deadlines and the six board questions
The executive summary singles out duties with fixed time limits: reporting a personal data breach to the PDPC office within 72 hours of becoming aware under section 37(4), notifying data subjects where the risk is high, and answering rights requests within the periods the law sets. It also says high-risk projects, such as automated decisions affecting individual rights or training on sensitive data, should pass a data protection impact assessment (DPIA) before starting.
A one-page graphic lists six questions for boards and senior executives to ask before approving an AI project. In summary: is the organisation controller or processor, and does the vendor contract restrict use of the data, for example by banning training on it; what is the legal basis, and are the LIA or consent records complete; does the project need a DPIA; is there human oversight with real power to reverse the system; is there a channel for data subject requests; and are there logs and an incident plan that allow notification within 72 hours.
What was not read for this article
Not every chapter was read line by line. The following are described by their table-of-contents titles only: chapter 1 (principles and objectives), sections 4.2 and 4.3 on data minimisation and collection from sources other than the data subject, chapter 6 (risk assessment and management), chapter 7 (security measures, split into organisational, technical and physical), chapter 8 (managing personal data breaches in AI systems), section 9.1, Appendix A (the checklist for developing AI), and the glossary and reference appendices. Readers who build or procure AI systems should read those in the original.
What this means for Thai marketers
This section is analysis by Relevant Audience. The guideline is written for all sectors, and in the pages read it does not discuss advertising platforms by name. It does touch marketing directly in four places: an example in section 2.3 of a marketing team feeding customers' purchase history and website behaviour into an AI system to predict interests and offer personalised promotions, which the guideline treats as organisational use subject to the PDPA, the chatbot consent example in section 4.6, the direct marketing and profiling passage in section 9.3, and FAQ item 3.3, which says organisations should tell users from the start that they are talking to an automated system.
Applied to common marketing work, the sections above suggest some checks.
- Chatbots on websites and LINE official accounts. Conversation logs are likely to be personal data under section 4.1. If transcripts feed model training, section 4.6 points to a separate, unbundled choice. Teams running chat as part of LINE marketing should check what the bot discloses and what the vendor keeps.
- Customer data in AI tools. FAQ item 4.1 rates putting internally held personal data into a public generative AI service as high risk and suggests a policy against it or an enterprise version backed by a data protection agreement. Appendix B recommends an acceptable use policy and staff training.
- Audience modelling and personalisation. Predictive scoring and lookalike modelling resemble the profiling that sections 9.3 and FAQ item 6.1 mention. Whether a given campaign reaches the "significant effect" level in section 5.4 is a legal question the sections read do not answer for advertising.
- Privacy notices. Section 4.4 says notices should be AI-aware and distinguish using data to produce an answer from using it to train or fine-tune a model. A review of consent and tracking alongside a Google Analytics 4 setup is a natural moment to update that wording.
Because the guideline is non-binding, none of this creates a new obligation by itself. The underlying PDPA duties already exist, and the guideline sets out the practice the Committee recommends for meeting them with AI.
Questions about the PDPC AI guidelines
Are the PDPC AI guidelines legally binding?
No. The PDPC states that the guideline is preliminary advice, has no legal force, is not binding in the consideration of an offence and is not an interpretation or ruling on the PDPA. The PDPA itself and its subordinate rules remain binding.
When were the guidelines published?
The Committee's notification is dated 22 October 2026 and was published in the Royal Gazette on 2 November 2026. The PDPC office publicised it on its website on 16 November 2026.
Is a prompt typed into an AI chatbot personal data?
It can be. Section 4.1 lists prompts, user context, telemetry logs and embeddings as data that may remain personal data when they can be linked back to an identifiable person.
Can a company use existing customer data to train an AI model?
Possibly, with conditions. Section 5.1 and FAQ item 2.1 say this may be a new purpose beyond the one originally notified, so the organisation should check the PDPA section 21 exceptions and otherwise notify the new purpose and obtain consent, or anonymise the data first.
Do companies have to tell people they are talking to an AI chatbot?
The guideline says they should. FAQ item 3.3 advises telling users from the start that they are communicating with an automated system, under the transparency principle, so they understand its limits and risks.
The short version
On 16 November 2026 the PDPC publicised a non-binding reference, 81 pages in the gazette file, for anyone in Thailand who feeds personal data into AI. It answers practical questions about prompts, vendors, consent and deletion with section numbers a legal team can check. Relevant Audience builds marketing programmes for Thai brands, including AI search visibility work. For decisions about your own data, speak to a qualified Thai privacy lawyer.







