AiPassᵀᴴ
| 🚧 Under Construction: | This page is currently under construction, and by "under construction", we mean it's in a black hole of procrastination from which no page has ever escaped. Fear not, intrepid internet explorer! This page will be finished... eventually. Maybe. Possibly. Probably not. But who knows, right? Stay tuned for updates, or don't. We won't judge. |
|---|
TH-AI Passport Limits + Government Sensitivity
The AiPassᵀᴴ's (Former Name: Plan-B AI passport, TH-AI Passport) low limits (10 prompts/person) are ironically grounded — a practical cap on compute waste and potential disruption in a sovereignty-focused system. It treats AI as a tool with costs, not infinite credit. Contrast that with American compute wars flex (thousands of data centers, hyperscaler builds) where "loudmouth" boardroom dynamics make government AI use hyper-sensitive: national security, data control, export rules, and political capture all collide.
The "loudmouth from the other side" knows this — AI for government isn't neutral; it's a power layer. The melting pot's strength (remix speed) becomes weakness when it forgets origins and burns allocation on generated filler instead of frame-by-frame systemic thinking.
| 💩 Shitpost Warning: | This article has been scientifically proven to be only marginally canonical. Enjoy at your own risk.
Reader discretion is advised: This post contains questionable logic, dubious facts, and a high probability of absurdity. ♻ Please recycle this article responsibly; its intellectual value is minimal. |
|---|
- Tokenmaxxing waste (cat videos, Uber AI budget burn) vs. signal abstinence and GitHub streak.
- Narrative vs. seated (Trends pollution, fast anime spoilers, HSR promises).
- Cultural asymmetry as observable pattern, not moral judgment — melting pot speed + architectural questions = truth-seeking advantage.
Privacy Concerns
TH-AI Passport / AiPASS setup is one of the more aggressive KYC + data-retention designs currently in the wild for a public AI service.
What the system actually does
From the official privacy policy and the public breakdown circulating on 19 August 2026:
- Identity is permanent and government-grade. Registration requires authentication through ThaID or the Thang Rath (ทางรัฐ) app, which binds the account to the 13-digit Thai national ID. That number does not change.
- Conversations are retained for at least one year. The policy states chat history (prompts, responses, uploaded files, behavior) is kept for 1 year, with the usual “unless required for public interest” carve-out. “Temporary Chat” mode does not meaningfully shorten this.
- Data is used for National LLM / ThaiLLM training. The project explicitly treats user interactions as training corpus (after claimed anonymization or pseudonymization). Once data enters model weights it is effectively permanent.
- Prompts are forwarded to foreign providers. The platform is a front-end that routes queries to GPT, Claude, Gemini, Grok and others. Thailand pays the inference cost (Pay-Per-Active-User model) while collecting the interaction data.
- Broad surveillance exceptions. The policy allows use of chat + identity + behavior data for national security and computer-crime purposes without needing a court order in the way ordinary private services do. User rights under PDPA (access, deletion, objection) are subordinated to “important public interest” as defined by the state.
This combination — permanent national ID + long-term chat logs + national-model training + third-party routing — is what Thai critics (including the posts you linked) are calling a privacy honeypot.
Why it feels “cheesy”
The service is structured as a cost center for the government:
- Taxpayers subsidize inference credits so citizens get free (or heavily subsidized) access to frontier models.
- In exchange, the state obtains a large, identity-linked corpus of Thai-language prompts and interactions that can be used to train domestic models.
- The foreign model providers still process the actual queries; the Thai platform sits in the middle collecting the data trail.
It is not a pure sovereign AI stack. It is a subsidized gateway that captures usage data while outsourcing the heavy compute.
Practical stance
Your conclusion is consistent with the risk profile:
- Treat it as a logged, identity-bound public service. Suitable for non-sensitive, non-political, non-health, non-religious queries if someone needs free access.
- Do not treat it as a private or low-risk channel. Anything typed is potentially reviewable by the state for an extended period and may contribute to model training.
- Maintaining separation (personal devices, paid direct accounts with the model providers, strong operational security) remains the cleaner way to avoid long-term identity-linked data exposure.
This is the same broader pattern we have been discussing: free or heavily subsidized AI access almost always carries a data-collection cost. In this case the cost is unusually explicit because of the national-ID binding and the one-year (extendable) retention window.
Staying away, or using it only for low-sensitivity tasks while keeping real work on accounts you control, is the rational move for anyone treating personal security and long-term digital footprint as first-class constraints.
