AI Regulation Tracker  /  Legislation in force

Vietnam's AI Law now orders generative-AI providers to watermark output and disclose training data

Under Law No. 134/2025/QH15, in force since March 1, 2026, providers of generative AI must mark AI output as machine-readable, tell users when they are dealing with a machine, and disclose the data used to train the model. The training-data duty puts copyright exposure squarely on builders. A widely cited implementing decree number, reported as "Decree 142/2026," rests on a single secondary source and is unconfirmed against the official gazette.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

Vietnam's Law on Artificial Intelligence, No. 134/2025/QH15, has been in force since March 1, 2026, and its generative-AI provisions reach directly into how AI products are built and shipped. Providers of generative AI that produces text, image, audio, or video are directed to do three things: apply a machine-readable digital watermark to the content their systems generate, disclose the data sources used to train the model, and ensure users can tell they are interacting with AI rather than a person. Read together, these are duties on the builder, not options.

The three duties

The watermarking duty is the most concrete for engineering teams. The law points toward a machine-readable mark on AI-generated output, the kind a downstream system or platform can detect, rather than a visible logo alone. The disclosure-to-users duty addresses a different risk: a person should not be misled into thinking a machine is a human. For a chatbot or voice agent, that means a clear notice that the counterpart is AI.

The third duty is the one that stands out. Providers are directed to disclose the data sources used to train the model. Secondary analyses from Vietnam Briefing and OneTrust read this as a duty rather than a mere right of inspection, and they connect it to copyright exposure: if a builder must state what went into training, the provenance of that data becomes a live compliance question rather than a private matter. That framing is what makes Vietnam's approach worth a professional's attention.

For most product teams these obligations are less about a single feature and more about a habit. The user-disclosure duty is straightforward to satisfy once, but it has to hold across every surface a customer touches, including embedded chat, voice, and any synthetic-media output. The watermarking duty travels with the content itself, which means it needs to survive export, download, and reuse rather than living only inside the app. And the training-data duty is retrospective as much as forward-looking: a firm cannot document sources it never tracked, so the practical work starts with keeping records of what data trained which model version.

What the law does not settle

The statute sets the obligations. It does not, on its face, hand builders a finished technical specification. The operational detail, how a watermark must be encoded, what format a training-data disclosure takes, and where carve-outs apply, is largely delegated to an implementing decree. As of this writing, that decree's number is not reliably confirmed. A single secondary source has reported a "Decree 142/2026," but that number has not been verified against Vietnam's official gazette (vanban.chinhphu.vn), and it should not be treated as settled. The safer reading is that the duties flow from the AI Law's generative-AI provisions and that the implementing detail is still pending publication.

That gap matters for planning. A firm can and should act on the direction of travel now, output marking, AI disclosure, and training-data records, without over-committing to a specific technical standard that the decree may define differently.

Why a US or EU builder should care

For a US reader, the training-data duty is the sharp edge. It is more aggressive than the current US posture, where no federal law compels generative-AI providers to disclose their training sources. It tracks the European Union more closely: the EU AI Act's general-purpose-AI training-content summary requirement directs providers of general-purpose AI models to publish a sufficiently detailed summary of the content used for training. Vietnam is moving in the same direction, and doing so for generative AI specifically.

Practically, a US or EU vendor selling generative-AI tools into Vietnam inherits these obligations for that market. A company already preparing an EU AI Act training-content summary will find the Vietnamese duty familiar and can reuse much of the underlying work. A company that has done neither now has two jurisdictions, not one, asking the same question about what its models learned from.

The bottom line for now

Treat the three duties as live. Build output watermarking and AI disclosure into generative features aimed at Vietnamese users, and start assembling a defensible account of your training-data sources. Then watch the gazette: the implementing decree, whatever its final number, will convert these directions into testable technical requirements, and the firms that documented their data early will adapt fastest.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.