Vietnam's AI Law and Decree 142/2026/ND-CP: what businesses need to do

Legal Updates · October 2026

Decree 142/2026/ND-CP, in force since 1 May 2026, implements Vietnam's Law on Artificial Intelligence No. 134/2025/QH15 and puts its core obligations into practice: risk classification, conformity assessment, and marking and labeling of AI-generated content. We summarize the key points for businesses and the questions still open.

In roughly a year, Vietnam has gone from having no dedicated AI legislation to having a standalone law and its first core implementing decree. The Law on Artificial Intelligence No. 134/2025/QH15 was passed by the National Assembly on 10 December 2025 and took effect on 1 March 2026. On 30 April 2026, the Government issued Decree No. 142/2026/ND-CP detailing parts of the Law and measures for its implementation, effective 1 May 2026. A few days earlier, on 25 April 2026, the Prime Minister issued Official Dispatch No. 36/CD-TTg calling for stricter discipline in drafting and issuing the documents that implement new laws, a signal that the remaining guidance will be pushed forward.

The framework is not finished. Several pieces still depend on further instruments, including the list of high-risk AI systems to be issued by the Prime Minister, the rules on administrative penalties and the operating rules of the National AI Development Fund. Even so, many obligations already apply, and companies using AI in products or operations in Vietnam should start their review now.

Modeled on the EU AI Act, but more flexible

Vietnam's AI Law borrows heavily from the EU AI Act: risk-based classification, tiered obligations, transparency requirements and coverage of foreign organizations and individuals. The difference is flexibility. The high-risk list is issued by the Prime Minister, so it can be updated without amending the law. Sensitive applications can be tested in a regulatory sandbox. And the Law does not set fines itself; it leaves administrative penalties to the Government, so businesses need to watch that instrument to estimate their financial exposure.

What Decree 142/2026/ND-CP covers

Its main subjects are:

  • An AI one-stop portal and a national AI database
  • Three-level risk classification and conformity assessment
  • Transparency, labeling and watermarking of AI-generated content
  • Reporting and handling of serious incidents (a preliminary report through the one-stop portal within 72 hours for urgent cases, or 5 working days otherwise)
  • Regulatory sandboxes and financial support for testing
  • The responsibilities of each ministry in regulating AI

Key points for businesses

Self-classification, self-responsibility. Providers assess their own AI systems against the Decree's criteria, decide the risk level (high, medium or low) and comply with the matching obligations. This is faster than prior certification, but the company carries the risk of getting the classification wrong.

High-risk systems need conformity assessment before deployment. Providers may self-assess (with a technical file, taking legal responsibility for the result) or use a conformity assessment body; systems on the list requiring certification before use must be assessed by an independent, registered body. Conformity information must be published on the one-stop portal before use, so the file has to be ready before launch.

Trade secrets are respected. Risk classification files and transparency and explanation duties do not require disclosure of source code, model parameters, detailed algorithms, raw training data or trade secrets, unless another law requires it. For foreign technology companies weighing a Vietnam launch, this matters.

Marking and labeling of AI content already apply. Providers must mark AI-generated audio, images and video in a machine-readable format; deployers must give clear notice when publishing AI content that could mislead users and visibly label content that imitates a real person's appearance or voice or recreates real events. Systems already operating before 1 March 2026 have a transition period under Article 35 of the Law (18 months in health, education and finance; 12 months elsewhere); new systems must comply now. Chatbots and image, video or voice generation tools offered to users in Vietnam should be checked early.

A three-level sandbox. Levels 1 and 2 are approved by the provincial People's Committee or the relevant ministry, depending on the scope of testing. Level 3, for the most sensitive cases, such as those involving sensitive personal data or national security, requires approval from the Ministry of Public Security.

Foreign providers. Under the Law, a foreign provider of a high-risk AI system offered in Vietnam must have a lawful contact point in Vietnam; where the system is subject to mandatory conformity certification, it must have a commercial presence or an authorized representative in Vietnam.

The grey area: AI system or AI model?

The Law and the Decree are built mainly around the AI system. The Law mentions foundation models and general-purpose AI models but does not set clear, separate obligations for standalone models. That leaves open questions for foundation models and general-purpose AI (GPAI), AI delivered through APIs, fine-tuned models, and models embedded in several different products.

The question we hear most often is this: when a Vietnamese company fine-tunes a foreign provider's model and builds it into its own product, who is the provider, and how is responsibility split along the value chain? The current rules do not give a clear answer. Until guidance arrives, the parties should allocate it in their contracts: who classifies the risk, who keeps the technical file, who reports incidents, and who pays if a regulator orders remediation.

What to do now

  • Map every AI system in use or planned, including tools bought from third parties
  • Classify the risk level of each system and record the reasoning
  • Prepare technical documentation for systems likely to be high-risk
  • Check watermarking and labeling obligations for every user-facing generative AI feature
  • Cross-check AI risk assessments against the data protection impact assessment (DPIA) and cross-border transfer impact assessment (CTIA) required under Vietnam's personal data protection rules, to avoid duplicate work and reuse documents where possible
  • Review contracts with model and API providers for allocation of responsibility

For foreign-invested companies, this work should run at group level and in Vietnam at the same time. EU AI Act documentation can be partly reused, but it does not replace the file required under Vietnamese law.

How Bizlink can help

We help companies map and classify their AI systems, prepare conformity assessment and personal data files, and draft and review licensing agreements for AI models and services. As new implementing rules are issued, we brief clients on what changes for their sector.

This article is general information, not legal advice on any specific matter. Current as of October 2026.