Lessons from Other Countries’ Mistakes: Why No Country Has Yet Solved AI in School

Whenever a major reform is discussed, a temptation arises: let us not reinvent the wheel, let us see how advanced countries have done it and copy the best. In the case of artificial intelligence in schools, this temptation is especially strong. It seems that somewhere someone must already have found a working model, and all that remains is to adopt it.

The bad news is that there is still nothing to copy. The good news is that this opens opportunities.

An analysis of international experience leads to a sobering conclusion: no country in the world today has a complete and stable model for integrating AI into school education. Even those usually called leaders are moving through trials, pauses, and revisions. This does not mean that foreign experience is useless — on the contrary. It means that what should be extracted from it is not ready-made recipes, which do not exist, but lessons, especially from other countries’ mistakes. The loudest lesson of the past year came from a country that is difficult to suspect of technological backwardness.

It is worth clarifying at once who is usually called a leader. In the international literature, Singapore, Estonia, South Korea, several U.S. states, and China are most often mentioned. But on closer inspection, their “leadership” is partial and narrow. None of them has built a holistic system. Each is strong in one area and stumbles in another. That is why it is more useful to look not at their showcases, but at their bruises.

South Korea is the dream of any technocrat: advanced digital infrastructure, strong schooling, a culture of education, and readiness to invest enormous resources. If anyone should have succeeded, it was South Korea. That is why its failure is so revealing.

The government launched an extremely ambitious AI textbook program — digital learning materials in mathematics, English, and computer science that promised to personalize learning, reduce teachers’ workload, and lower dropout rates. Huge sums were spent: more than a trillion won of public money, plus hundreds of billions invested by publishers. The textbooks were made mandatory and introduced in schools at the start of the academic year in March 2025.

Four months later — after a single semester — everything collapsed. The textbooks were stripped of official status and reclassified as “supplementary materials,” meaning that schools themselves would decide whether to use them. More than half of the schools that had joined the program immediately abandoned it.

What went wrong? Almost everything skeptics had warned about. The textbooks were full of errors. Lessons were disrupted by technical failures — connection and authorization problems. Instead of reducing workload, teachers received additional work for which they had not been properly prepared. Parents protested, fearing increased screen time and digital dependence among children. Teachers’ unions and parent associations filed lawsuits, arguing that implementation had been made mandatory while risks and teachers’ own views were ignored. A change of government also played a role: the incoming administration treated the previous government’s project skeptically.

The bitter irony is that one of the program’s declared goals was to reduce inequality by decreasing Korean families’ dependence on expensive private tutoring centers. In practice, the rushed launch risked adding burdens to both children and teachers. Another sobering detail: the idea of using AI to “generate textbooks faster than usual” meant that development, review, and preparation were given three to four times less time than ordinary printed textbooks require. When children are involved, rushing verification is costly.

The main Korean lesson is not that “AI in school does not work.” It is that the technology was not destroyed by technology. It was destroyed by haste, compulsion, and disregard for those who work in the classroom. Top-down implementation without preparing teachers and without their consent dooms even a well-funded project. A teacher must see AI as an ally, not as an obligation imposed by order.

At the other pole is Singapore, often cited as an example of a careful, institutionally calibrated approach. The difference from Korea is not that Singapore implements AI more slowly, but that it implements it differently.

In Singaporean schools, AI is used primarily as a tool for analytics and support: analyzing learning data, helping plan instruction, and experimenting with adaptive systems that diagnose a student’s knowledge gaps in real time. Specific examples show the logic. An adaptive mathematics learning system uses machine-learning algorithms to identify where exactly a particular student is struggling during the lesson itself. An English-language assistant takes over routine checking of written work — grammar, syntax, and spelling — relieving teachers of the mechanical part. But — and this is crucial — official documents explicitly emphasize that AI does not replace the teacher but serves as an auxiliary tool. Most solutions are introduced not across the whole country at once, but through limited pilots: first test on a small scale, make sure it works, and only then expand.

This sounds less spectacular than “AI textbooks for everyone from Monday.” But this caution is itself a strategy. Singapore is not trying to jump into the future in one leap; it is moving toward it in steps, each of which can be tested and, if necessary, reversed. In this respect, it is closer to a reasonable model than many louder reformers.

Estonia is known as one of the world leaders in digital government, where almost everything from voting to taxes has long worked online. In education, it decided not to be overly cautious, but to take a large step — though a step built differently from Korea’s.

In 2025, Estonia launched the national AI Leap program, deliberately echoing the legendary Tiger Leap program of the 1990s, which brought computers and the internet into schools thirty years ago and made the country a digital pioneer. From September 1, 2025, the first wave — around 20,000 upper-secondary students and 3,000 teachers — received free access to leading AI tools for learning, with plans to expand coverage further to colleges. Major global AI developers became partners. In scale, this is not a timid pilot but a national move.

But the key point is that Estonia placed boldness inside an institutional framework — exactly the framework Korea lacked. First, a special fund involving the state and the private sector was created to manage the program; the initiative has a responsible institutional carrier, not just a ministry order. Second, teacher training was planned not “later,” but in advance, before tools were given to children; the teacher is made a co-author rather than someone trying to catch up. Third, the goal is explicitly framed not as “AI instead of people,” but as “AI will strengthen learning”: as one of the program’s architects put it, the winner is not the one who uses AI most, but the one who uses it most intelligently. AI is built not as a fashionable separate subject, but as a cross-cutting skill: students use it to generate ideas, debug code, and analyze sources while learning to treat algorithmic output critically.

The Estonian lesson is therefore not simply “do not rush,” but something more subtle: one can move boldly and quickly, but only if boldness rests on an institution, trained teachers, and an honest orientation toward strengthening rather than replacing the human being. Estonia and Korea both made a large bet. The difference is that one built support beneath it, while the other relied on the force of a mandatory order.

The two largest technological powers show how mistakes can be made from opposite sides.

The United States is an example of innovation from below. There is no single national policy; decisions are made by districts and individual schools. On the one hand, this produces strong local cases and rapid experimentation. On the other, it creates chaos and unpredictability. New York’s story is telling: at the beginning of 2023, fearing that students would cheat, the country’s largest school district banned ChatGPT. A few months later the ban was lifted and the approach shifted toward regulated access, after it became clear that a blunt block does not work and merely pushes use into the shadows. Moving from ban to admission within a single year is the visible price of lacking a thought-through framework. The cost of this fragmentation is a high risk of inequality between wealthy and poor schools, and legal uncertainty.

China represents the opposite extreme: strict control from above. The country conducts large-scale experiments with AI analytics and personalization, while also keeping the technology on a short leash. During the national university entrance exam — gaokao — in June 2025, major AI services temporarily disabled photo recognition in their chatbots to prevent students from scanning assignments. At the same time, AI is used to monitor behavior during those exams. The country is also engaged in difficult debates about control, data, and psychological impact on children, while regulatory restrictions have only intensified in recent years.

Two paths — maximum freedom and maximum control — both run into their own dead ends. The answer, as usual, is not at the poles.

If these cases are compared, it becomes clear that countries with very different systems are stepping on the same rakes. This may be the main observation of the entire international section.

The first is social resistance. Parents and teachers across the world — in Europe and Asia alike — voice the same concerns: weakening of independent thinking, increased screen time, and opacity of algorithms. This is not the caprice of isolated conservatives, but a stable social reaction.

The second is growing inequality. Digital and AI solutions, as international organizations note, often do not reduce but increase the gap between schools with different levels of resources, staff preparedness, and infrastructure. A technology meant to equalize opportunities often strengthens stratification in practice.

The third is regulatory deadlock. Questions about who owns educational data, who is responsible for an algorithm’s error, and how to reconcile a person’s right to delete their data with the need to store a long educational history remain unresolved even in developed legal systems.

The fourth is revision and rollback. Again and again, large-scale initiatives — whether AI textbooks or automatic assessment — return to pilot mode or are abandoned altogether after a wave of criticism.

The conclusion from this set of rakes is clear: most problems of implementing AI in education are institutional, not technological. The problem is not that algorithms are insufficiently smart. The problem is that society, school, and the state are not managing to build rules, trust, and responsibility. This means the solution is not the purchase of technologies, but the construction of institutions.

There is one more pattern visible when international documents are viewed in dynamics. Over the past few years, the very tone of the conversation about AI in education has changed.

Early strategies, roughly from 2016 to 2019, breathed techno-optimism: AI would arrive, sharply increase the efficiency of learning, personalize everything, and solve long-standing school problems. Later documents from leading international organizations sound noticeably different. In them, AI is increasingly described as a factor that complicates governance of the education system rather than simplifying it. The slogan “implement as widely as possible” is being replaced by a logic of staged movement, protected experiments, mandatory evaluation of effects, and — this is essential — the possibility of stopping.

Serious analysis speaks in the same vein. A major international publication, examining the boldest predictions about AI transforming schools, directly described them as based on an oversimplified understanding of what education is. School does not merely transmit knowledge. It forms discipline, communication skills, responsibility, and the ability to think independently over years. These things are hard to automate, and anyone promising to “replace school with an algorithm” most likely does not understand how learning actually happens — through effort, error, live interaction, and feedback.

This turn from enthusiasm to sobriety may be the most mature outcome of global experience. It is not a rejection of AI, but a rejection of illusions about it.

It is easy to draw a pessimistic conclusion from all of this: if even the leaders have not succeeded, perhaps we should not try. The right conclusion is the opposite.

The absence of universal models means that Kazakhstan is in the same conditions as the rest of the world. No one has moved far ahead; there is no ready-made model to chase. This is not backwardness — it is an open window. In this situation, the winner is not the one who buys and copies someone else’s technology fastest, as we have seen where that leads, but the one who builds its own governance of the process more intelligently.

International experience does not give Kazakhstan recipes, but it clearly marks the minefield. It says: do not implement from the top down; do not ignore teachers; do not make untested solutions mandatory; prepare infrastructure before platforms; build rules for providers; protect data and national context; and preserve the ability to slow down. This is the basis on which a Kazakhstani trajectory can be built — from understanding, not copying.

There is one more lesson, the most uncomfortable one. Before bringing AI into school, it is worth looking honestly at the school itself and at its unresolved problems. Experience shows that AI may not cure these problems, but amplify them. This is the subject of the next, perhaps sharpest, material in the series.

The material was prepared based on the results of foresight sessions held in autumn 2025 and winter 2026.