What is Vietnam's AI Law 2025 and who does it apply to?
Law No. 134/2025/QH15 on Artificial Intelligence was passed by the 15th National Assembly on 10 December 2025 and took effect on 1 March 2026 (Article 34). It has 8 chapters and 35 articles and governs the research, development, provision, deployment and use of AI systems in Vietnam; AI used exclusively for national defence, security and cipher work is out of scope (Article 1).
Its reach is broad: it applies to Vietnamese agencies, organisations and individuals and to foreign organisations and individuals taking part in AI activities in Vietnam (Article 2). A company that simply buys an off-the-shelf AI tool to serve customers can therefore carry obligations, not only companies that build models.
As of October 2026, two further instruments matter: Decree 142/2026/ND-CP (issued 30 April 2026, effective 1 May 2026), which details the Law, and Prime Minister's Decision 33/2026/QD-TTg (signed 30 June 2026, effective 15 August 2026), which issues the List of high-risk AI systems. This guide cites mainly the Law itself; points from the decree and decision are attributed to their sources.
Four roles to identify first
Obligations attach to roles, not to company size. Article 3 defines:
- Developer: designs, builds, trains, tests or fine-tunes all or part of an AI model, algorithm or system and directly controls the technical methods, training data or model parameters (clause 3).
- Provider: places an AI system on the market or puts it into use under its own name or brand, whether it built the system itself or a third party did (clause 4).
- Deployer: uses an AI system under its control in professional, commercial or service activities; personal, non-commercial use is excluded (clause 5).
- User: interacts directly with an AI system or uses its output (clause 6).
Core principle: humans stay in control and keep responsibility
Article 4 sets four principles. The one that matters most to system builders is clause 2: AI serves people and does not replace human authority or responsibility; human control and the ability to intervene must be maintained over all decisions and behaviour of an AI system, together with system safety, data security and the ability to inspect and supervise how the system is developed and operated.
The other clauses require a human-centred approach that protects privacy (clause 1); fairness, transparency, non-discrimination and accountability for AI decisions and consequences (clause 3); and green, energy-efficient AI (clause 4).
In engineering terms, this becomes concrete requirements: an approval step before an AI takes a consequential action, a stop or override mechanism, and logs that show who decided what. It is the same thinking behind the human-in-the-loop pattern many teams already use.
Risk classification: which tier is your AI system in?
Chapter II (Articles 9–15) is the part businesses should read most closely. Article 9 sorts AI systems into three risk tiers; classification considers the impact on human rights, safety and security, the sector of use (especially essential sectors) and the number of users and scale of impact (Article 9(2)).
| Risk tier | Definition (Article 9) | Main obligations | Supervision (Article 10(5)) |
|---|---|---|---|
| High | May cause significant harm to life, health, legitimate rights and interests, national or public interests, or national security | Classification dossier, notification to the Ministry of Science and Technology, conformity assessment before use (Art. 13), risk management, technical file, logs, human oversight (Art. 14) | Periodic inspection or inspection on signs of violation |
| Medium | May confuse, influence or manipulate users because they cannot tell they are dealing with AI or with AI-generated content | Classification dossier, notification to the Ministry, transparency under Art. 11, explanation on request (Art. 15(1)) | Reports, sample checks or independent evaluation |
| Low | Anything not in the two tiers above | Explanation when there are signs of violation or impact on rights (Art. 15(2)); encouraged to publish basic information | Monitoring upon incidents or complaints |
Who classifies, and when notification is required
The provider self-classifies before putting a system into use; medium- and high-risk systems need a classification dossier (Article 10(1)) and the provider must notify the result to the Ministry of Science and Technology through the national AI single-window portal before use (Article 10(3)). If the tier is unclear, the provider may ask the Ministry for guidance based on its technical file (Article 10(4)).
A deployer may rely on the provider's classification, but if it modifies, integrates or changes functions in a way that creates new or higher risk, it must work with the provider to reclassify (Article 10(2)). This is easy to miss: once you connect a language model to customer data and let it make decisions on its own, you may have built a system in a different tier from the original product.
The high-risk list is now in force
Decision 33/2026/QD-TTg places 46 AI systems in the high-risk group across six sectors: education; ethnic and religious affairs; health; banking; judicial proceedings; and transport. Education examples include AI that automatically tests, grades or ranks learners and AI that monitors and analyses learner behaviour; in banking, AI that automatically decides on credit.
If your product touches these activities, read the annex of Decision 33/2026/QD-TTg itself rather than relying on summaries.
Obligations of providers and deployers of high-risk AI
Article 14 separates responsibilities between the parties. The table helps a business see what it must do as provider or as deployer.
| Provider (Article 14(1)) | Deployer (Article 14(2)) |
|---|---|
| Establish, maintain and review risk-management measures when the system changes significantly (point a) | Operate and monitor the system within its classified purpose, scope and risk level without creating new risk (point a) |
| Govern training, testing and operational data for quality (point b) | Ensure data security and human intervention during use (point b) |
| Prepare, update and retain the technical file and activity logs (point c) | Keep complying with AI standards and technical regulations in operation (point c) |
| Design the system for human oversight and intervention (point d) | Meet transparency and incident duties under Articles 11 and 12 (point d) |
| Transparency and incident handling; explain to authorities and inform users and affected persons (points dd, e) | Explain operation, risk control and incident handling; inform users and affected persons (point dd) |
| Cooperate in inspection, post-market review and incident remediation (point g) | Cooperate with the provider and authorities (point e) |
Other points worth noting
- Explanations do not require disclosure of source code, detailed algorithms, parameter sets or trade secrets (Article 14(1)(e); Article 15(1)(b)).
- High-risk systems need a conformity assessment before use or after a significant change; systems on the mandatory-certification list are assessed by a registered body, while others may be self-assessed by the provider or assessed by a registered body (Article 13).
- Foreign providers of high-risk systems offered in Vietnam must have a lawful point of contact in Vietnam (Article 14(6)).
- If a high-risk system is operated correctly but still causes harm, the deployer compensates the injured party first and may then claim reimbursement from the provider or developer if the parties have agreed so (Article 29(2)). Reimbursement clauses in AI procurement contracts therefore matter.
Transparency and labelling of AI-generated content
Article 11 also applies to medium-risk systems, so it is the obligation most businesses running chatbots, virtual assistants or content tools will meet first.
- Providers must design AI systems that interact directly with people so that users know they are interacting with AI (clause 1).
- Providers must ensure AI-generated audio, images and video are marked in a machine-readable format (clause 2).
- Deployers must clearly disclose when they publish AI-generated or AI-edited text, audio, images or video that could mislead people about the authenticity of events or persons (clause 3).
- Deployers must label, in an easily recognisable way, AI audio, images or video that imitate a real person's appearance or voice or re-create real events; cinematic and artistic works are labelled in a suitable manner (clause 4).
- Both must maintain transparency information throughout the time the system or content is provided (clause 5).
What Decree 142/2026/ND-CP adds
According to an analysis by law firm Atsumi & Sakai, Article 18 of Decree 142 requires deployers to apply visible labels to AI-generated or AI-edited audio, images and video that imitate a real person's appearance or voice or re-create real events, and exempts certain cases such as technical editing, text processing, internal use, and research and testing (Article 18(4)). A practical review workflow is covered in our guide to labelling AI-generated content.
Prohibited acts and serious incidents
Article 7 lists six groups of prohibited acts. For businesses, three stand out:
- Using fake or simulated people or events to deliberately and systematically deceive or manipulate people in a way that causes serious harm (clause 2(b)); exploiting the vulnerabilities of vulnerable groups such as children, the elderly or persons with disabilities (clause 2(c)).
- Collecting, processing or using data to develop, train, test or operate AI in breach of laws on data, personal data protection, intellectual property and cybersecurity (clause 3).
- Obstructing, disabling or falsifying human oversight, intervention and control mechanisms (clause 4); concealing information that must be disclosed, or erasing or falsifying mandatory labels and warnings (clause 5).
Incident handling
When a serious incident occurs, developers and providers must urgently fix, suspend or recall the system and notify the competent authority; deployers and users must record and promptly report the incident and cooperate in remediation (Article 12(2)). According to Atsumi & Sakai, Article 19 of Decree 142 sets a preliminary reporting deadline of 72 hours for emergency and loss-of-control incidents and 5 working days for other serious incidents.
During inspections, organisations must provide technical files, trace logs, training data and other necessary information (Article 28(3)). Without logs, it is very hard to show a system was run properly.
Transition deadlines: when must existing AI systems comply?
For AI systems already in operation before 1 March 2026, Article 35 gives providers and deployers 18 months from the effective date for systems in health, education and finance, and 12 months for all other sectors. During this period systems may keep running unless the authority finds a risk of serious harm (Article 35(2)).
For systems on the high-risk list that were operating before 15 August 2026, Decision 33/2026/QD-TTg sets deadlines of 1 September 2027 for health, education and finance and 1 March 2027 for the other listed sectors (per official publications of the Decision).
| Date | What happens |
|---|---|
| 10 Dec 2025 | National Assembly passes the Law on Artificial Intelligence |
| 1 Mar 2026 | Law takes effect (Article 34) |
| 1 May 2026 | Decree 142/2026/ND-CP takes effect |
| 15 Aug 2026 | List of high-risk AI systems (Decision 33/2026/QD-TTg) takes effect |
| 1 Mar 2027 | End of the 12-month period under Article 35 for sectors other than health, education and finance; deadline for listed high-risk systems in other sectors |
| 1 Sep 2027 | End of the 18-month period for health, education and finance; deadline for listed high-risk systems in these sectors |
Technical checklist: what to prepare now
The Law sets requirements at the level of principle; the real work sits in system architecture and operating procedures. The list below is EYC's suggestion based on the articles cited and does not replace advice from a lawyer.
1. Inventory and classification
- List every AI system in use: chatbots, internal assistants, AI Agents, content generators, scoring models. Record your role for each (provider or deployer).
- Self-assess the risk tier under Article 9 and check it against the list in Decision 33/2026/QD-TTg; keep your reasoning on file.
- Flag integrations or customisations that may raise risk above the original product (Article 10(2)).
2. Logging and traceability
- Log inputs, outputs, model version, tools called and the approver for consequential decisions (supporting Article 14(1)(c) and Article 28(3)).
- Give AI Agents scoped permissions and keep audit logs; see access control and audit logs for AI Agents.
- Set log retention periods consistent with personal data law.
3. Human control
- Add human approval before consequential actions: sending information externally, changing records, making decisions about people.
- Provide pause, override and rollback, and test regularly that they work (Article 4(2), Article 7(4)).
- Name an accountable owner for each system.
4. Transparency for users
- Tell customers clearly when they are talking to AI (Article 11(1)).
- Label AI content that imitates real people or real events; preserve machine-readable marking metadata where your tools support it (Article 11(2), (4)).
- Prepare user-facing documentation describing functions, how the system works and risk warnings.
5. Documentation, data and contracts
- Keep a technical file: intended purpose, main data sources, test results, risk controls.
- Check the legal basis for data used by AI under personal data protection law; see Vietnam's Personal Data Protection Law 2025 and data governance basics.
- Add to contracts with AI vendors: their classification result, support for explanations, incident notification and reimbursement of compensation (Article 29(2)).
- Build an incident response procedure with an owner, a report template and internal deadlines shorter than the legal ones.
Where EYC can help
EYC is a technology and AI company based in Ho Chi Minh City. Through our integration and consulting service, we can work with your team to inventory AI workflows, design logging, permissions and human approval points, and prepare technical documentation as input for your legal team. Scope and responsibilities are set under contract. EYC does not provide legal advice; conclusions about legal obligations should come from your lawyers.
EYC also builds Early Years Circle, a platform for preschools. Because education has its own transition deadline and includes systems on the high-risk list, we follow these rules closely as we develop the product.
Frequently asked questions
When does Vietnam's AI Law take effect?
Law 134/2025/QH15 took effect on 1 March 2026 (Article 34). AI systems already running before that date have 18 months to comply in health, education and finance and 12 months in other sectors (Article 35).
Does a company that only uses ChatGPT or other off-the-shelf AI tools need to comply?
It may. A business using an AI system in professional, commercial or service activities is a deployer under Article 3(5) and has duties such as disclosure and labelling of AI content (Article 11(3), (4)) and explanation on request (Article 15). If you put your own brand on an AI system offered to customers, you may become a provider.
Which AI systems count as high-risk in Vietnam?
Article 9 defines high-risk systems as those that may cause significant harm to life, health, legitimate rights and interests, or national and public interests. The concrete list is in Decision 33/2026/QD-TTg: 46 systems across six sectors—education, ethnic and religious affairs, health, banking, judicial proceedings and transport.
Must AI-generated content be labelled in Vietnam?
Under Article 11, deployers must clearly disclose AI content published to the public that could mislead about authenticity, and must visibly label audio, images or video that imitate real people or re-create real events. Providers must mark AI-generated audio, images and video in a machine-readable format. Decree 142/2026/ND-CP sets out details and exemptions.
What are the penalties for violating the AI Law?
Depending on the nature and severity, organisations and individuals may face administrative penalties or criminal liability and must compensate for any damage (Article 29(1)). Administrative penalty levels are set by the Government (Article 29(5)).
Which authority oversees AI in Vietnam?
The Government has overall responsibility; the Ministry of Science and Technology is the national focal point; other ministries coordinate, and provincial People's Committees manage AI locally (Article 30(2)).