The Vietnam AI law is Law No. 134/2025/QH15 on Artificial Intelligence. The National Assembly passed it on 10 December 2025 and it has applied since 1 March 2026. It sorts AI systems into three risk tiers (high, medium, low) and puts most duties on the "provider", meaning whoever places an AI system on the market or into use under its own name or brand. The "deployer", the business that uses AI in its operations, owes transparency, labelling, oversight and incident-reporting duties. The details sit in Decree 142/2026/ND-CP (in force 1 May 2026) and in the list of high-risk AI systems attached to Decision 33/2026/QD-TTg (in force 15 August 2026).
For most foreign-invested companies the practical answer is reassuring but not empty. A customer-service chatbot, a CV-screening tool or a credit-scoring model with a human credit officer in the loop is not on the high-risk list. You still need to work out whether you are a provider or a deployer; if you are the provider, you self-classify each system and notify the Ministry of Science and Technology (MOST) when a system is medium risk. You also need to label synthetic media of real people and comply with Vietnam's Personal Data Protection Law. Systems already running before 1 March 2026 have a 12-month transition (to around 1 March 2027), or 18 months (to around 1 September 2027) in healthcare, education and finance (Article 35 of Law 134/2025/QH15).
Vietnam AI law timeline#
| Date | Instrument | What it means |
|---|---|---|
| 10 Dec 2025 | Law 134/2025/QH15 passed | Risk tiers, prohibited acts, transparency, liability |
| 1 Mar 2026 | Law in force (Article 34) | Chapter IV (AI) of Law 71/2025/QH15 on Digital Technology Industry repealed (Article 33) |
| 1 May 2026 | Decree 142/2026/ND-CP in force (Article 45) | Classification criteria, dossiers, notification, labelling, incidents, sandbox |
| 15 Aug 2026 | Decision 33/2026/QD-TTg in force (Article 3) | 46 high-risk AI systems in 6 sectors |
| 28 Aug 2026 | MOST announces the national one-stop AI portal (ai.dda.gov.vn) is live | Channel for classification notices and incident reports |
| 1 Mar 2027 | 12-month transition ends (Article 35; see below on the exact last day) | Systems in use before 1 March 2026, outside health, education, finance |
| 1 Sep 2027 | 18-month transition ends (Article 35) | Systems in health, education and finance |
The law excludes AI used solely for defence, security and cryptography (Article 1(2)), and it expressly covers foreign organisations and individuals taking part in AI activities in Vietnam (Article 2). A parent company that pushes a group AI tool to its Vietnamese subsidiary can therefore fall within scope in its own right.
Who the Vietnam AI law regulates: four roles#
Article 3 of the Vietnam AI law defines the roles that drive every obligation:
- Developer (clause 3): designs, trains, tests or fine-tunes a model or system and directly controls the technical method, training data or parameters.
- Provider (clause 4): places an AI system on the market or into use under its own name, brand or trademark, whether it built the system or a third party did.
- Deployer (clause 5): uses an AI system under its control in professional, commercial or service activities.
- User (clause 6): interacts directly with the system or uses its output.
The law also protects "affected persons" (clause 7), such as a rejected job applicant or a loan applicant turned down on the basis of an AI output.
The provider definition is the one FDI companies tend to miss. If your Vietnamese entity puts a chatbot built on a foreign model API on its website as "[Brand] Assistant", it has put an AI system into use under its own brand. On our reading it is both the provider and the deployer of that chatbot, and the model vendor does not take over the classification duty for your system. The table below is also our reading of common situations; there is no official guidance on them yet.
| Situation | Your role |
|---|---|
| Staff use a licensed GenAI assistant for internal drafting | Deployer; staff are users |
| Branded website or messaging chatbot running on a third-party API | Provider and deployer |
| Off-the-shelf SaaS CV-screening tool, used as supplied | Deployer |
| Model fine-tuned on your data and sold to customers | Developer and provider |
Risk tiers under the Vietnam AI law and the high-risk list#
Article 9 of the law sets three tiers. A system is high risk only if it appears on the Prime Minister's list (Article 6(3)(a) of Decree 142/2026/ND-CP). It is medium risk if users could be confused, influenced or manipulated because they cannot tell they are dealing with AI or AI-generated content (Article 9(1) of the Decree). Everything else is low risk.
The provider self-classifies before putting the system into use and answers for the result (Article 10(1) of the Law; Article 6 of the Decree). A deployer inherits that classification but must reclassify with the provider if its own changes, integrations or new purposes create new or higher risk (Article 10(2) of the Law). Classification applies to systems, not to stand-alone models (Article 6(2) of the Decree).
Article 9(3) of Decree 142/2026/ND-CP takes five situations out of the medium tier. The two that matter most for businesses are office tools whose AI nature is obvious from context and that cannot impersonate anyone, and systems that do not interact with, or deliver services or content to, the public (directly or through a platform).
Decision 33/2026/QD-TTg lists 46 high-risk AI systems: education (3), ethnic and religious affairs (7), healthcare (2), banking (2), judicial proceedings (1) and transport (31). Each entry carries an "applies only when" condition. For example, AI-driven credit decisions are high risk only if the system makes the final approval or rejection, the decision is used to lend or disburse without an independent credit officer's approval, and the loan meets a value threshold set by the State Bank of Vietnam. Education entries cover official automated grading and biometric monitoring of learners, among others. Surgical robots, autonomous vehicles, traffic and air-traffic control, aviation security and real-time biometric identification at transport hubs make up most of the rest. There is no entry for recruitment, HR, marketing or customer service. The two banking entries allow provider self-assessment of conformity under Article 13(2)(b) of the Law rather than prior certification.
Vietnam AI law obligations by role and risk tier#
Transparency (Article 11) and incident handling (Article 12) apply to any system with the relevant features. The remaining duties depend on tier:
| Tier | Provider | Deployer | User |
|---|---|---|---|
| High | Classification dossier; notify MOST before use; conformity assessment published on the portal; risk management, data governance, technical file and logs; human oversight by design; accountability to regulators (Articles 10, 13, 14 of the Law; Articles 12–15 of the Decree) | Operate within the classified purpose and risk; data security and human intervention; accountability; liability to compensate even when the system was used correctly (Article 29(2)) (Article 14(2) of the Law; Article 15(4)–(6) of the Decree) | Follow operating procedures, no unauthorised changes, report incidents to the deployer (Article 14(3)) |
| Medium | Classification dossier; notify MOST through the portal before use; tell users they are dealing with AI; machine-readable marking of AI audio, images and video; explain on request (Articles 10, 11, 15(1) of the Law; Articles 12, 14, 17 of the Decree) | Disclose and label AI content released to the public; explain on request (Article 15(1)(c) of the Law; Article 18 of the Decree) | Follow disclosure and labelling rules |
| Low | No notification; publishing basic information is encouraged; explain if there are signs of a violation (Articles 10(3), 15(2)) | Explain if there are signs of a violation | Lawful use at own responsibility |
Points worth knowing:
- The classification dossier covers identification, a functional description and context of use, a summary of the main input data, and risk controls. Source code, weights and raw training data do not have to be disclosed, and the dossier is kept for the life of the system (Article 12(2), (4) and (6) of the Decree).
- Notification is a self-declaration, not a licence. The portal issues a system ID and an electronic acknowledgement on submission, and the authority may not ask for extra documents at that stage (Article 14(4) of the Decree).
- An upward reclassification must be notified within 15 working days of completing the review (Article 11(3)(a) of the Decree).
- A foreign provider of a high-risk system offered in Vietnam needs a lawful point of contact in Vietnam; if the system requires prior certification, it needs a commercial presence or an authorised representative (Article 14(6) of the Law). This is the provision most likely to reach overseas group companies directly.
AI labelling and transparency#
Article 11 of the Vietnam AI law and Articles 16–18 of Decree 142/2026/ND-CP require four things:
- Chatbots and virtual assistants must be designed so that people know they are interacting with AI (Article 11(1)).
- Providers mark AI-generated audio, images and video in a machine-readable format, for example through file metadata or a digital signature; text is exempt unless another law says otherwise (Article 17(1)–(2) of the Decree). For open-source or free tools, a deployer that publishes the output must apply the marking itself (Article 17(5)).
- Deployers must clearly disclose AI-generated or AI-edited content released to the public if it could mislead people about real events or persons (Article 11(3)).
- AI labelling is mandatory for audio, images and video that simulate a real person's face or voice or recreate real events (Article 11(4) of the Law; Article 18(2) of the Decree). An advertisement using a cloned likeness of a celebrity needs a label even if the celebrity consented.
Labels must be clear and shown before or when the audience first sees the content. They can sit on the content, in the caption, in the platform interface, or be spoken (Article 18(3) and (5)). No label is needed for pure quality enhancement, for spell-checking, summarising or translating text without changing its meaning, for internal-only content, or for closed testing (Article 18(4)). The Decree gives no separate grace period for labelling, and Article 35 of the Law only covers systems already running before 1 March 2026. One could argue that content made with a tool in use before that date is still within the transition, but on our cautious reading new AI-made marketing content should carry labels now.
Article 7 lists the prohibited acts, including deceptive deepfakes that cause serious harm, exploiting children and other vulnerable groups, using data in breach of data protection, IP or cybersecurity law to train or run AI, disabling human oversight, and removing mandatory labels. Unlike the EU AI Act, the Vietnam AI law has no separate "unacceptable risk" tier; banned uses are written as prohibited acts.
Serious incident reporting#
A serious incident is one where an AI system causes death or serious injury, significant property loss, a serious infringement of rights, or serious disruption of essential services, security or public order (Article 19(1) of the Decree). The provider or the deployer files a preliminary report through the portal (form AI01a for organisations):
- within 72 hours of confirming the incident for harm to life or health, disruption of essential services or security, and uncontrollable infringements of rights;
- within 5 working days for other serious incidents (Article 19(3)).
The clock starts once there is enough initial information to show the incident is real and probably caused by the AI system; filing is not an admission of fault. If the deployer cannot reach the provider, the deployer files. A final report on remediation follows within 15 days of the preliminary report (Article 19(4)). Where the same event is also a personal data breach, the 72-hour notice from discovery under Article 23(1) of Law 91/2025/QH15 may apply as well. Article 19(5) of the Decree says that where an incident must also be reported under cybersecurity, data protection or sector law, reporting follows that law, and authorities should not ask again for information they already hold. The Decree does not say whether form AI01a must still be filed on the AI portal in that case, so prepare for both channels (see the 72-hour data breach rule).
Which authority enforces the Vietnam AI law#
Under the AI law in Vietnam, the Government holds overall authority and MOST is the lead ministry; line ministries regulate AI in their sectors and provincial People's Committees act locally (Article 30(2) of the Law). Under Decree 142/2026/ND-CP:
| Task | Authority |
|---|---|
| Classification notices, conformity results, incident reports via the portal; drafting the high-risk list | MOST (Articles 4, 14, 19, 41) |
| Sector-specific safety rules, inspections, sandbox levels 1 and 2 across two or more provinces | Line ministries such as the Ministry of Health or the State Bank of Vietnam (Article 43) |
| Sandbox levels 1 and 2 within one province; local inspection | Provincial People's Committees (Article 44) |
| Sandbox level 3; incidents affecting national security or public order | Ministry of Public Security (Articles 23(3)(c), 42(5)) |
When an AI system processes personal data, the A05 Department, Ministry of Public Security, remains the data protection authority under Law 91/2025/QH15. The two tracks run independently. A voluntary regulatory sandbox is also available (Articles 21–27 of the Decree), with decisions by the provincial People's Committee, the line ministry or the Ministry of Public Security in 10, 20 or 30 working days respectively (Article 23(5)).
Penalties under the Vietnam AI law and other liability#
Article 29 of the Vietnam AI law provides for administrative fines, criminal liability and civil damages, and leaves administrative penalties to a Government decree. The law itself contains no fine amounts. As of 13 October 2026 we have not found a decree in the Official Gazette that sets fines for breaches of the AI law, such as failing to notify a classification or to label content.
Enforcement already exists through other routes:
- Regulators can order reclassification, extra documents or suspension of a system that was misdeclared (Article 10(6)), and suspension or recall after a serious incident (Article 12(3)).
- Decree 330/2026/ND-CP is the penalty decree for cybersecurity and personal data protection, not for the AI law's own duties; it bites only where an AI system processes personal data. Its Article 67 fines organisations VND 20–50 million for skipping the annual data protection compliance assessment; VND 50–70 million for failing to disclose or explain how an automated algorithm works, to offer an opt-out from automated processing, or to classify the AI system by risk for data protection purposes; and VND 70–100 million for automated decisions affecting people's rights without an oversight mechanism or a right to human review, with a possible additional 3–6 month suspension of the system. Individuals pay half (Article 7(1)).
- For high-risk systems used correctly that still cause damage, the deployer compensates the victim and can recover from the provider or developer only if the contract allows it (Article 29(2) of the Law). Check your vendor contracts for that recourse clause.
Transition deadlines under the Vietnam AI law#
Providers and deployers of systems in operation before 1 March 2026 have 18 months from the law's effective date in health, education and finance and 12 months elsewhere to comply, and the systems may keep running unless the regulator orders suspension or termination for serious risk (Article 35). Systems launched on or after 1 March 2026 get no transition under Article 35 (but see the high-risk rule below).
On the exact last day: under Articles 147 and 148 of the Civil Code, 12 months from 1 March 2026 ends at the close of 1 March 2027 (18 months: 1 September 2027), but Article 4 of Decision 33/2026/QD-TTg says "before 1 March 2027" and "before 1 September 2027". On our cautious reading, finish by 28 February 2027 (or 31 August 2027).
For listed high-risk systems, Article 4 of the Decision sets those two deadlines for systems in use before 15 August 2026, and "before 1 March 2027" for systems launched within six months after that date. If a system has to be reclassified as high risk, Article 11(5) of the Decree allows up to 12 months from the date the amended list takes effect, provided real human oversight and logs are kept. The clause cross-refers to significant functional changes (Article 11(1)(a)) while counting time from a list amendment; we read it as covering list amendments, but confirm with MOST for a specific case.
The Gazette version of Decree 142/2026/ND-CP sets no separate deadline (such as 60 days) for notifying systems already in use; its transitional Article 46 only covers the period before the portal went live. The deadlines for existing systems are those in Article 35 of the Law and Article 4 of Decision 33/2026/QD-TTg.
Where the Vietnam AI law overlaps with the PDPL#
- Training data: training AI on data obtained in breach of data protection law is a prohibited act under Article 7(3) of the AI law. Decree 356/2025/ND-CP allows personal data to be used for AI development if the data rules are met, and treats inferences that identify a person as personal data (Article 10(1)–(2)). See de-identification.
- Automated decisions: controllers must tell data subjects about automated processing, explain the logic and its effect on their rights, and offer an opt-out (Article 10(3) of Decree 356/2025/ND-CP). AI processing of personal data must also be risk-classified under Article 30(4) of Law 91/2025/QH15, a separate exercise from AI-law classification.
- Biometrics: biometric data is sensitive personal data (Article 4(1)(dd) of Decree 356/2025/ND-CP) with specific security duties (Article 31(4) of Law 91/2025/QH15). Face-recognition time clocks are not high-risk AI, but these duties still apply.
- One dossier for two laws: Article 12(7) of Decree 142/2026/ND-CP lets a provider use its DPIA as, or as part of, the AI risk classification dossier.
- Offshore model APIs: sending prompts that contain personal data to servers abroad is a cross-border transfer under Article 20 of Law 91/2025/QH15.
Teams that built their AI governance around the GDPR should check the differences in GDPR vs Vietnam's PDPL rather than assume the same rules apply. Primary texts: Law 91/2025/QH15 and Decree 356/2025/ND-CP.
Three common cases#
Branded customer chatbot. You are provider and deployer. Because the chatbot talks to the public, it does not fit the Article 9(3) exclusions and, on our reading, is likely medium risk: prepare a classification dossier, notify through the portal before launch (or before your transition deadline if it ran before 1 March 2026), show a clear "you are chatting with AI" notice and a route to a human, and document the cross-border transfer if the model is hosted abroad.
Third-party CV screening. You are the deployer. Recruitment is not on the high-risk list and a tool that only ranks CVs for HR does not face the public, so on our reading it is low risk under the AI law. The heavier duties are in data protection law: candidate data may be used only for recruitment, needs consent and must be deleted if the candidate is not hired unless otherwise agreed (Article 25(1) of Law 91/2025/QH15), and candidates must be told about and given an explanation of automated processing, with an option to opt out (Article 10(3) of Decree 356/2025/ND-CP). Missing those mechanisms can cost VND 50–70 million (Article 67(2) of Decree 330/2026/ND-CP); if the tool rejects candidates automatically without oversight or a right to human review, the range is VND 70–100 million (Article 67(3)(b)).
Credit scoring. High risk only if the system makes the final decision without an independent credit officer and the loan meets the State Bank threshold (item IV.2 of the Appendix to Decision 33/2026/QD-TTg). Keep a credit officer with real power to override, and the system stays off the list. Finance systems in use before 1 March 2026 have the 18-month transition, and customers' credit information is sensitive personal data (Article 4(1)(k) of Decree 356/2025/ND-CP), so a DPIA is still needed.
Vietnam AI law compliance checklist for FDI companies#
- Inventory every AI system you use or offer in Vietnam, with vendor, purpose, inputs, affected people and go-live date.
- Assign a role to each system, paying attention to anything branded with your name.
- Check each system against the Decision 33/2026/QD-TTg list, including the "applies only when" conditions.
- For public-facing systems, decide medium risk or exclusion; prepare the dossier, reuse your DPIA, and notify via ai.dda.gov.vn where you are the provider.
- Add AI disclosure and a human hand-off to chatbots.
- Issue an AI labelling rule for marketing: synthetic faces and voices of real people are labelled, and tool watermarks are never stripped.
- Merge AI incident reporting (72 hours, 5 working days, 15 days) with your personal data breach process.
- Amend vendor contracts: technical information for accountability (Article 16(6) of the Decree), reclassification support, incident cooperation and recourse for compensation.
- Close the data protection gaps: automated-processing notices, opt-out, annual compliance assessment, cross-border transfer dossier.
- Adopt an internal AI use policy and diarise reviews ahead of the March 2027 and September 2027 deadlines, and whenever the high-risk list changes.
Much of the AI law in Vietnam runs on the same documents as data protection compliance: the system inventory, the DPIA, the incident playbook and vendor contracts. LexData reviews AI systems from the personal data angle, prepares DPIAs that describe the AI system, and builds incident procedures as part of our PDPL compliance consulting.
