On 21 August 2026, opposition leader Natthaphong Ruengpanyawut of the People's Party held a Bangkok press conference and presented documents he said show a surveillance risk in the TH-AI Passport. The 1.6-billion-baht programme offers free premium AI tools to up to 5 million users. His warning: users could be identified and their prompts monitored by state agencies or service providers. The Digital Economy and Society Ministry replies that data is stored in Thailand, is accessible only in anonymised form, and that contracts bar AI providers from retaining it or training on it. Public access opened on 31 August.
The programme itself is worth defending. Giving millions of Thais free access to frontier AI tools is exactly the kind of diffusion that raises productivity, and the ministry reports that more than 1.13 million people had pre-registered by 21 August. The question is not whether to do it, but under what rules.
The strongest case for the government
A state that pays for a service has legitimate reasons to log it. A free scheme open to everyone aged 15 and over invites abuse: generating illegal content, scams, or automated bulk use that burns public money. The ministry says it screens and masks personal or sensitive data before prompts reach foreign providers, and that providers may not retain data or use it to train models (Thairath). Keeping records in Thailand is also a reasonable response to data-sovereignty concerns, and the ministry says this follows the Personal Data Protection Act and cybersecurity laws.
If those assurances were the whole story, the opposition's alarm would be overstated.
What the privacy notice reportedly says
They are not the whole story. Thai outlet Naewna, reviewing the programme's privacy notice, reported that item 13 provides for storing AI conversation data and the content users enter, with the stated purpose of monitoring inappropriate use (Naewna). According to that report, if an algorithm flags content as possibly violating the Computer Crime Act or affecting national security, a human officer can open the account history and read the chats, suspend the account, and pass the data to law enforcement. Conversation data is reportedly kept for one year from each message, and item 12 covers cookie-based tracking of usage behaviour under consent. We could not independently access the notice itself, so treat these details as reported rather than confirmed.
Taken at face value, this does not match the picture of data accessible only in anonymised form. Anonymised aggregates may serve dashboards, but a flagged-chat review pathway is by definition identifiable. The ministry's two statements can both be true only if the notice distinguishes routine analytics from an exception for flagged content. That exception is the thing that deserves scrutiny.
Why the exception matters
"National security" and Computer Crime Act triggers are broad categories, and AI prompts are unusually revealing. People use chatbots to draft legal worries, health questions, political opinions and half-formed ideas. A user who knows an officer might read a flagged conversation has reason to self-censor, which is a speech cost even if no one ever reads their chat.
Thailand's data protection framework offers less comfort than it might. Law-firm analysis by Tilleke & Gibbins describes a 2023 royal decree under the PDPA that partially exempts data controllers from certain requirements when specified state agencies request personal data, and lets controllers disclose without consent when a legally authorised agency cites its statutory power (Tilleke & Gibbins). Security duties remain, but the disclosure channel is open. Section 4 of the Act itself also carves out certain state activities such as national security and public safety; the statute text is available in unofficial English translations (PDPA 2019, Mahidol University copy). One academic reading of those carve-outs argues the Act gives little guidance on safeguards that exempted authorities must follow.
The result is that a user's main protection is the ministry's contract and notice, not an independent check.
Procurement questions compound the trust problem
The scheme has also drawn procurement criticism. The People's Party argued in June that terms of reference changed capacity figures and compressed timelines during a caretaker period, and Prime Minister Anutin Charnvirakul said the project would proceed (Nation Thailand). The ministry later revised the contract to a pay-per-active-user model, with the Attorney General's office reviewing it (Thairath). A payment model tied to active users also creates an incentive to log and measure behaviour, which makes data-minimisation commitments more important, not less.
A proportionate fix
None of this requires cancelling the programme. Four changes would address the real risk without sacrificing the benefit:
- Publish the privacy notice and the data-processing contracts, including the flagging criteria and who counts as an authorised officer.
- Require a court order or an independent authority's approval before identifiable chat content is handed to law enforcement.
- Shorten retention from one year to the minimum needed for abuse prevention, and delete content once a flag is cleared.
- Commission an independent audit by the Personal Data Protection Committee, with results made public.
The lesson reaches beyond Thailand. Governments that subsidise AI access become custodians of very intimate data. The test of a good programme is not where servers sit, but who can read a citizen's prompt, under what legal standard, and with what oversight. Answer those in law, and TH-AI Passport can be both widely used and trusted.