Việt Nam Names Its High-Risk AI: 46 Systems Now Under the Strictest Rules

Six months ago today, on 1 March 2026, Việt Nam’s first dedicated AI law took effect. It promised a risk-based regime — but a risk-based regime is only as real as its risk list. That list now exists and is legally in force. Quyết định 33/2026/QĐ-TTg, signed by Deputy Prime Minister Hồ Quốc Dũng on 30 June 2026 and effective since 15 August 2026, names the 46 AI systems the government classifies as high-risk, spread across six sectors. If you build, sell, or deploy any of them in Việt Nam, your compliance clock is already running.

What the document is

Quyết định 33/2026/QĐ-TTg is a Prime Minister’s decision issuing the Danh mục hệ thống trí tuệ nhân tạo có rủi ro cao — the catalogue of high-risk AI systems. It is the third and most concrete layer of a regulatory stack that came together remarkably quickly:

Luật Trí tuệ nhân tạo số 134/2025/QH15, passed by the National Assembly on 10 December 2025 and in force since 1 March 2026 — Việt Nam’s first standalone AI law. It classifies AI systems into three risk tiers (high, medium, low) and defines high-risk systems as those capable of causing serious harm to life, health, lawful rights and interests, public interests, or national security.
Nghị định 142/2026/NĐ-CP, issued 30 April 2026 and effective 1 May 2026 — the implementing decree. It sets up the national AI portal and the national database of AI systems under the Bộ Khoa học và Công nghệ, requires providers to declare medium- and high-risk systems through the portal, mandates conformity assessment for high-risk systems before deployment, and requires machine-readable markers on AI-generated audio, images, and video.
Quyết định 33/2026/QĐ-TTg — the catalogue that tells everyone which systems the “high-risk” obligations actually bite on.

What is on the list

The 46 systems fall into six sectors, and the distribution is lopsided in a telling way. Transportation dominates with 31 of the 46 entries: high-autonomy vehicle control, road and rail traffic management, railway dispatching, automated traffic-violation detection, and systems managing critical transport infrastructure. The other five sectors are narrower but pointed:

Healthcare — AI that assists in surgery and robotic surgical systems that autonomously perform treatment actions on a patient’s body.
Banking — systems that execute high-value transactions automatically, and AI that approves or rejects credit without a human officer’s review.
Education — AI that serves self-study content from uncontrolled data sources, and systems that autonomously grade learners or assess their psychology and behaviour.
Litigation — biometric identification systems deployed at scale for investigating and resolving cases.
Ethnic affairs and religion — systems that automatically score applications, verify eligibility, or allocate policy support without human approval.

The common thread is autonomy over consequential decisions: the list targets AI that acts on bodies, money, verdicts, benefits, and moving vehicles without a human in the loop.

What changes compared to before

Before 15 August, the AI Law’s high-risk tier was a definition without an address. Providers could argue about whether their system met the abstract harm threshold; regulators had no enumerated basis for enforcement. The catalogue converts that ambiguity into a closed list — a notably different technique from the EU AI Act, which defines high-risk by broad use-case categories and leaves classification partly to the provider’s own assessment. Việt Nam’s approach trades flexibility for certainty: you are either on the list or you are not, and the government can amend the list as technology moves.

Being listed carries the full weight of the Nghị định 142/2026 obligations: conformity assessment before the system goes live, declaration in the national AI database, a documented risk-management system, guaranteed human oversight and the ability to intervene, and incident reporting. The decision also restates the operating principles — a high-risk system cannot change or replace the legal responsibility of the agencies and officials it assists, and it must remain subject to human supervision, control, and intervention at all times.

Who it affects and the deadlines

For systems already operating before 15 August 2026, the decision sets staggered transition deadlines — and, interestingly, the sectors with the longest lists get the shortest runway. Systems in transportation, litigation, and ethnic affairs must comply by 1 March 2027 — six months from today. Systems in healthcare, education, and finance get until 1 September 2027. Systems put into operation within six months after the effective date also fall under the 1 March 2027 deadline. During the transition, an existing system may keep running unless authorities determine it poses a risk of serious harm, in which case it can be suspended.

The practical impact lands first on Việt Nam’s transport-tech sector — autonomous driving pilots, smart-traffic vendors, railway operators — and on banks running automated credit models. Edtech companies using automated assessment, and hospitals procuring surgical robots, have a year longer but a heavier assessment burden. Foreign providers are not exempt: the AI Law and Nghị định 142/2026 apply to foreign organisations providing AI systems into Việt Nam.

Open questions

The list’s precision is also its weakness: 46 named systems will not stay current for long, and how quickly the catalogue gets amended — and whether additions come with their own transition periods — will determine how predictable the regime feels. It is also not yet clear how conformity assessment will work in practice: who the assessment bodies are, what the technical standards will require, and how a transport operator with a dozen listed systems is meant to sequence certification before March. And the boundary cases will be litigated in guidance, not text — is a bank’s credit model “autonomous” if a human rubber-stamps 99% of its outputs? The first amendment cycle and the first enforcement actions will tell us more than the decision itself does.

Sources

This post is general information, not legal advice.