A field report from inside Vietnam’s AI curriculum: 35,000 teachers, 170 pilot schools, and the free app students actually use
By: Robert Sterling – AsiaPresswire – The room smelled of chalk dust and overheating laptops. It was the last period on a Tuesday at a lower-secondary school in District 12, Ho Chi Minh City, and the Grade 8 English teacher had just finished showing her class how to rewrite a paragraph by asking a chatbot to “make it more formal.” The students nodded along the way students nod when they are not entirely sure whether what they are doing counts as schoolwork or as something else entirely. On the whiteboard, the lesson objective was written in Vietnamese: “Using AI to support learning.” On the projector, the chatbot’s cursor blinked, waiting for the next prompt.
This was not one of the 170 pilot schools. The 170 — the ones the Ministry of Education and Training had designated for the AI curriculum pilot beginning in March 2026 — came with their own materials, their own schedules, their own evaluators. This school had none of those. What it had was a teacher who had learned on her own, during the pandemic and after, that a free account with a well-known American chatbot could produce a passable lesson plan, a set of comprehension questions, and, when the class grew restless, a joke in Vietnamese that bought thirty seconds of silence.
The ministry, to its credit, knows this is the majority case. A survey of nearly 35,000 teachers earlier this year found that 76% had already used AI in the classroom. A separate survey of more than 11,000 students found 87% were aware of AI applications in education. Those are not small numbers. They describe a country where artificial intelligence arrived without a procurement budget, without a textbook, and without a single server on Vietnamese soil.
This article is about what happens next.
Classroom Time Written into Law Across Southeast Asia
On 18 August 2026, the Ministry of Education and Training issued Decision 2422/QD-BGD_DT, and the next day a companion circular, 5588/BGD_DT-GDPT. Together they set out the national AI education framework for primary and secondary schools, to take effect from the 2026–2027 school year. The headline figure is the one that will appear in every summary: twelve lessons per class, per year, of core AI content. Add supplementary material and the total is higher, but the twelve is the hard, quantified, auditable number.
Vietnam is, by this measure, the only country in Southeast Asia to write a quantified AI lessons target into a national decision. Thailand has ten pilot smart classrooms under a UNESCO–Huawei programme. Malaysia has eleven AI-enabled classrooms in trial and a stated plan for eight hundred, though the eight hundred has not appeared in any official document anyone has published. The Philippines has a forty-nine-page AI guideline from the Department of Education, the first K-12 national AI framework in the region, but it specifies behaviour, not hours. Indonesia is prioritising English teacher training and has made AI an optional junior-high subject. Vietnam has legislated the clock time.
The framework itself is built around four knowledge strands and three levels. The strands are human-centred thinking, AI ethics, AI techniques and applications, and AI system design. The levels map onto the school system:
| Level | Grade band | What students do |
| Level 1 — Familiarisation | Primary | Understand basic concepts |
| Level 2 — Foundation | Lower secondary | Develop ethics, technique and application basics; begin analysing AI’s social impact |
| Level 3 — Creation | Upper secondary | Use AI with confidence; understand basic principles; design and fine-tune simple AI tools; debate social issues, human rights and civic responsibility |
The three modes of delivery are: a dedicated AI lesson plan slotted into the existing double-period timetable; integration into existing subjects without changing or adding to the subject learning goals; and optional activities — clubs, discussions, field visits, extracurricular work — for the supplementary content.
And then there are the three protections, which are the part of the decision that says the most. Students are not required to hold a personal AI account. Students may not use a personal account or an unvetted AI tool. Schools are responsible for screening and vetting AI tools against the relevant grade band, educational goals and terms of use, and for supervising and guiding students while they use them. No other government in the region has written any of this into a national document. The first clause, in particular, is an admission: the state has looked at what children are actually doing and concluded that letting them each bring their own account into the classroom is beyond its control.
A Curriculum Written After the Practice
The decision is best understood as a ratification, not an introduction. The practice ran two years ahead of the legislation. In early 2024 the ministry surveyed more than 11,000 students across the provinces and found 87% already aware of AI applications in education. In early 2025 it surveyed nearly 35,000 teachers and found 76% had already used AI in the classroom. By the time Decision 2422 was signed in August 2026, Vietnamese teachers had been using chatbots for at least three school years, and the official state news agency reported — in its own words — that “most teachers and students use free online applications such as ChatGPT and Gemini.”
The timeline tells the story plainly:
- 2024: the ministry issues its first guidance on AI teaching and learning.
- 2025: it instructs the development of an AI education content framework.
- February 2026: the first pilot training course is held.
- August 2026: Decision 2422 sets twelve lessons per class per year, to take effect from the 2026–2027 school year, after a year of piloting.
The legislation landed after the practice had already settled. That is not a criticism of the ministry. It is a description of the only way a large, slow bureaucracy can move in a country where the internet reaches the classroom faster than any procurement process. But it does mean that the curriculum is, in effect, formalising a habit the state did not create and cannot fully control.
The framework’s own background survey lists the obstacles, and they are worth taking at face value: AI content is still mostly folded into information technology and STEM activities rather than standing as an official unified course; there is a lack of teachers with deep AI teaching expertise, most of whom have learned on their own; and students’ ability to use AI tools varies widely. These three difficulties were published by the state news agency. They are not leaks. They are the ministry’s own assessment.
Voices From the Teaching Floor
The survey numbers form a chain, and each link is weaker than the one before it. Eighty-seven percent of students are aware of AI in education. Seventy-six percent of teachers have used it in class. Just under 70% of those teachers hold a positive view of it. Awareness, use and approval: each step down the ladder loses roughly ten points, and the drop from “have used” to “think it is a good idea” is the most revealing. A tool that three-quarters of the workforce has tried but that barely two-thirds endorse is a tool with unresolved problems.
The teachers’ stated worries are not about whether they can operate the software. They are about accuracy, about privacy and data security, and about the reduction of human-to-human interaction in the classroom. Those are not the anxieties of people who have not figured out the interface. They are the anxieties of people who have used it enough to know where it fails.
The same survey, and the ministry’s own reporting, identify three structural difficulties that sit behind those anxieties. First, there is no unified curriculum — AI content is still stitched into information technology and STEM activities rather than standing as its own subject with its own syllabus, its own materials and its own assessment. Second, there is a shortage of teachers with genuine depth in AI pedagogy; most have taught themselves. Third, students’ ability to use AI tools is uneven, and the gap between the confident and the confused is wide.
The decision addresses the first of these. It does not address the second or the third, because those are not problems a decision can solve. Teacher training is running in two phases across seventeen provinces, but the gap between “trained” and “competent in a subject that is itself moving” is a gap that no timetable closes quickly. And the third problem — uneven student ability — is a social problem that happens to show up inside a classroom, which means a curriculum can accommodate it but cannot fix it.
What the decision does do, concretely, is give a principal something to schedule and a ministry something to report. Each province must submit an implementation plan by 30 August each year and a report on results, difficulties and obstacles by 15 June each year. That is real accountability, of a kind that exists in few other countries in the region. Whether it is accountability that improves anything depends on what the reports say, and nobody outside the ministry has read them yet.
Domestic AI Law, Cybersecurity Law, and Data Leaving the Country
While the education system was formalising its relationship with American chatbots, the state was building the most comprehensive digital regulatory apparatus in Southeast Asia — and the two projects sit uneasily next to each other.
The AI Law took effect in March 2026. It is the first dedicated artificial intelligence law in Southeast Asia, and it follows a risk-based model modelled broadly on the European approach. AI systems are regulated across their full lifecycle — research, deployment, use — and high-risk AI systems must undergo rigorous safety assessment and certification before they can be used. The law also mandates data localisation for certain categories of data and requires foreign technology companies to maintain a physical presence or representative office in the country so that regulators can reach them.
The Cybersecurity Law, effective 1 July 2026, merges two earlier laws into one. It imposes strict liability for deepfakes, makes social media group administrators responsible for content, and establishes a comprehensive real-name system for online activity. The Personal Data Protection Law, effective 1 January 2026, requires a cross-border transfer filing within sixty days of the first transfer of personal data abroad, a seventy-two-hour breach notification, and a mandatory data protection officer for any organisation processing personal data. Small and micro enterprises are exempt; small and medium enterprises get a five-year grace period provided they do not handle sensitive data and do not exceed 100,000 records. A draft Data Security Law was submitted to the National Assembly in October 2026, aiming to set new controls on the cross-border movement of core and important data.
And in the same period, police detained eighty-six persons in connection with cyber fraud operations.
The ambition behind all of this is not confined to regulation. In Dien Bien, a mountainous province bordering Laos and China, the government opened a low-altitude economy sandbox in July 2026, running for eleven months through May 2027, with 6,000 test flights, a flight-safety target of 99%, a goal of validating three to five business models and cutting approval time and cost by 20 to 30%. The sandbox exempts participants from roughly sixty testing, certification and technical-compliance requirements — a number that quantifies, better than any policy speech, how heavy the regulatory burden is outside the sandbox. Nearly ten other provinces have expressed interest in following. The point is not the drones. The point is that the same state that is writing twelve AI lessons into a national decision is also writing a sandbox for an entirely new economic sector, and that both projects share a conviction that the way to manage technology is to legislate its boundaries in advance rather than to let the market draw them.
Read these together and the picture is strange. The state is constructing, in a single year, the densest data-protection and AI-regulation framework in the region — while the default tool in the country’s classrooms is a free American website that sends prompts and learning data to servers outside Vietnam, under terms of service no Vietnamese parent or principal has read. The PDP Law’s sixty-day cross-border filing requirement and the Cybersecurity Law’s real-name system are, in principle, exactly the tools to address that. In practice, enforcing them against a free consumer service used by millions of schoolchildren is a different proposition from filing paperwork for a corporate data transfer.
There is a sharper version of the same point. The AI Law’s risk-based framework and the education decision’s requirement that schools vet tools are, between them, a complete regulatory answer to the question “which AI may Vietnamese children use.” The answer the law provides is: one that has passed a safety assessment, is certified, is screened by the school, and complies with Vietnamese data-protection rules. The answer the classroom provides is: whatever is free and works. The gap between those two answers is the distance between the legislation and the reality, and it is measured not in years but in the number of servers that sit between a Vietnamese child’s homework prompt and the model that answers it.
Hardware Gaps: 170 Schools Meet a Free Website
Ho Chi Minh City launched its AI pilot in about 170 schools in March 2026, with an evaluation planned for June and a national rollout proposal to follow. That is a serious number by regional standards. Thailand’s entire UNESCO–Huawei smart-classroom programme covers ten schools. Malaysia has eleven in trial. Put the three side by side and Vietnam’s 170 looks like the leading edge of something.
But the denominator matters. Ho Chi Minh City alone has more than 7,000 schools. The 170 are roughly 2.4% of that. Nationwide the figure is smaller still, and the ministry’s own survey data suggests the practice is already far more widespread than the pilot, because the practice does not require a pilot at all — it requires a laptop and an internet connection.
The hardware layer across Southeast Asia is, in effect, being divided between two powers, and Vietnam’s share of that division is the passive one. In Thailand, the smart classrooms run on Huawei networks and interactive displays under a UNESCO partnership, with a parallel Chinese coding programme in two schools. In Malaysia, the government’s pilot partner list runs to Microsoft, Intel and Dell, with local education companies Pandai and Chumbaka alongside them. In the Philippines, the Department of Education’s approved tool list names Khanmigo, Canva AI and Microsoft Copilot, with ChatGPT and Gemini permitted subject to ethical and teaching standards. In Vietnam, the tools are ChatGPT and Gemini, chosen by teachers and students spontaneously, without procurement, without audit and without anyone having decided whether that is a good idea.
The education decision’s clause that students do not need a personal AI account is, read carefully, an acknowledgment of defeat on this point — the state is conceding that it cannot stop children from bringing their own accounts to class, so it will instead pretend the accounts do not exist and require the school to vet the tools instead. That is a pragmatic compromise, but it does not solve the underlying dependency. It merely moves the choice from the child’s pocket to the principal’s office, and the principal’s office has no more options than the child did.
The deeper constraint is the one the ministry’s own survey names: equipment and technical resources are insufficient. In the provinces, the obstacle is not pedagogical ambition. It is that a significant share of classrooms lack reliable electricity, reliable connectivity, or both. Cambodia provides the regional benchmark for how low this can go — only 22% of Cambodian schools have functional computers — and Vietnam sits well above that but well below the level at which a curriculum built around cloud-based AI tools can be delivered at scale. The twelve-lessons-a-year target is a policy instrument. It is not a description of what is physically possible in every classroom that is supposed to deliver it.
The Boom and Bust of Southeast Asian Edtech
The most instructive thing that happened in Southeast Asian education finance in the first quarter of 2026 was not the record. The region’s edtech startups raised roughly US$2.8 billion in the quarter, more than double the same period last year. The largest single transaction was KKR’s US$1.3 billion acquisition of a majority stake in XCL Education Holdings, a Singapore-incorporated school operator with campuses in Singapore, Bangkok and — relevant here — the Vietnam Australia International School in Ho Chi Minh City.
The transaction is not an edtech deal. XCL operates physical K-12 schools. Its revenue is tuition. Its assets are buildings, teachers and enrolment rolls. Counting it in edtech startup funding is a category error, and a revealing one: the only way to make the region’s edtech funding look like a record is to include a private-equity buyout of brick-and-mortar international schools. Strip that out and the underlying picture is considerably less cheerful.
That picture has been known for some time. Zenius, one of Indonesia’s oldest and best-known edtech companies, shut down entirely in January 2024 after more than two decades in business. Ruangguru, the region’s most prominent survivor, has not raised capital since its Series C in April 2021 — more than five years ago — having accumulated US$205 million across nine rounds. Its monthly visits fell from 19.4 million to 10.7 million between late 2024 and late 2025, a decline of 37.9%, and 97.1% of its traffic comes from Indonesia alone. The official user count is 15 million; a database estimate puts it at 40 million, which is a reminder that even the user numbers in this sector are unreliable. Google, Temasek and Bain’s e-Conomy SEA report — the authoritative scorecard for the region’s digital economy — stopped tracking edtech as a separate category after 2021.
The companies that survived did not survive by selling education to students. Ruangguru’s 2026 partnership list adds, in recent months, a state-owned mining group, a Singaporean bank and an ASEAN tourism association. It now works with thirty-two of Indonesia’s thirty-four provincial governments and more than three hundred twenty district governments, and it runs a financial-literacy programme for micro and small enterprise owners in Jakarta in partnership with a payment company and the city government. That is a company that used to sell K-12 tutoring. It is now, primarily, a vendor to governments and corporations.
Vietnam’s own bright spot in this space is Techcoop, which raised a US$70 million Series A in early 2025 — reportedly the largest A round in Southeast Asian agritech history. It is a farming-data and services company, not an edtech company, which is itself the point: in Vietnam, the most investable digital-native company working at the intersection of technology and livelihood is in agriculture, not education.
The backdrop against which all of this is happening is worth stating, because it is the reason the region’s attention keeps returning to the country. Vietnam’s gross domestic product grew between 7.63% and 8.18% in the first half of 2026, registered foreign direct investment rose 61% over the same period, and the country welcomed 12.4 million tourist arrivals. Those are numbers that describe an economy in unusually good health, and they sit behind the investor appetite that produced the KKR deal and the Techcoop round alike. But growth does not automatically produce a domestic edtech market. It produces the conditions in which an international school operator can sell tuition to a growing middle class, and in which a farming-services company can raise seventy million dollars — and it leaves the consumer education software business, in Vietnam as across the region, to be worked out by someone else.
The capital story and the policy story are running in opposite directions. The state is legislating more aggressively than any neighbour — twelve lessons a year, a dedicated AI law, a data-protection framework with teeth — while the private capital that might build the tools to deliver that curriculum has, in the region as a whole, been withdrawing from the consumer education market for years. The international capital that is moving into education in Vietnam is buying tuition-paying schools, not software. KKR’s portfolio in Asian education runs through Lighthouse Learning in India, Taylor’s in Southeast Asia, EQuest and Vinschools in Vietnam, Cognita in Asia and now XCL. Every one of those is a physical school. Not one is a software product.
What This Means for a Vietnamese Student in 2026
So what is actually happening in the classroom, and what is likely to happen over the next few years?
The honest answer is that both things are true at once, and the tension between them is the story. Vietnam is building the most ambitious state-led AI education framework in Southeast Asia — quantified, legislated, auditable, with a dedicated law, a data-protection regime and a national cybersecurity architecture behind it. And the default mechanism for delivering AI education to the country’s children is a free American website, chosen by teachers, used by students, and beyond the reach of any Vietnamese regulator.
The framework is real and worth taking seriously. Twelve lessons a year is more than any neighbour has committed to. The four-strand, three-level structure is coherent. The three protections — no personal account, no unvetted tools, school-level screening — are the right instincts, and they are written into a binding decision rather than a ministerial speech. The reporting cadence gives the policy a chance to be evaluated rather than announced.
But the framework has three weaknesses that no amount of legislative drafting will fix, because they are not legislative problems.
The first is the tool dependency. Until Vietnam or another country in the region produces an AI model and platform that teachers trust as much as ChatGPT, the curriculum will be delivered on someone else’s infrastructure, under someone else’s terms of service, with someone else’s data practices. The AI Law’s certification requirements and the data-protection law’s cross-border filing rules are the right tools for that problem, but they are tools aimed at enterprises. Applying them to a free consumer service used by schoolchildren is a enforcement problem that has not been solved anywhere in the region.
The second is the teacher gap. Seventy-six percent of teachers have used AI, but the survey’s own wording tells you what that means: they have used it, not been trained in it. The ministry’s pilot training covers seventeen provinces. The remaining several hundred have no structured pathway to competence in a subject that is itself changing faster than any curriculum cycle. The twelve-lessons-a-year target assumes a teacher who can deliver twelve lessons. It says nothing about where that teacher learned them or how confident they are doing so.
The third is the equipment gap, and it is the most concrete of the three. A cloud-based AI tool is only as available as the electricity and the connection. The ministry’s own survey names equipment and technical resources as a primary obstacle. Until that is solved, the twelve lessons are a policy target rather than a delivered reality in a significant share of the country’s schools.
None of this is a case for doing nothing. It is a case for being precise about what has been achieved. The legislation is an achievement. The survey data is an achievement — a country that knows how many of its teachers have used AI is ahead of most of the world. The pilot in 170 schools is an achievement, and the evaluation due in June will be the most useful document the ministry has produced on this subject, because it will describe what actually happened rather than what was planned.
The question to watch is not whether the twelve-lessons target is met. It is whether, by the time the 2027–2028 school year begins, there is a Vietnamese — or at least a regional — AI tool that a teacher in a provincial school can use without sending her students’ work to a server in California. If there is, the framework will look prescient. If there is not, the framework will still be in place, twelve lessons a year, delivered on someone else’s platform, and the children will still be using the free version — because that is what works, and because that is what the law, for all its ambition, has not yet found a way to change.