The State, Platforms, and Data: The Main Institutional Dilemma of AI in School

In the previous material, we reached a conclusion that changes the entire frame of the conversation: the problem of AI in school is institutional, not technological. Behind all the individual difficulties lies one hard knot that is the most difficult to untie. It is the question of who controls the data and digital platforms on which education now rests.

It sounds technical and dull — data, platforms, regulation. But behind this dry language lies perhaps the main story of the entire AI transformation. This is where it is decided who will actually govern the school of the future: the state, private companies, or no one at all, by default. Let us unpack this knot thread by thread.

Let us begin with an uncomfortable observation. Digitalization and elements of AI are already embedded in everyday school practice. But it is important to understand who makes the decisions that define this practice. The answer is discouraging: increasingly, decisions are made not by the state and not within educational strategies, but at the level of digital platforms that students, parents, and teachers use every day.

What does this mean in practice? The logic of learning processes, interfaces, available functions, and the mechanics of how a child studies and how a teacher works are increasingly set from outside — by those who created the platform. Often this happens outside the state system of education governance. School, in this picture, ceases to be an autonomous institution and becomes part of a large digital ecosystem — national and often global — whose rules were not written by the school.

This leads to the first fundamental conclusion. AI in education stops being a pedagogical question of “how to teach better” and becomes an institutional task of “how the architecture of the platforms on which everything happens is arranged.” If that is the case, it must be governed not at the level of a single lesson, but at the level of that architecture. Whoever designs it determines the future of school.

To feel how serious this is, one simple thought experiment is enough. Imagine that a platform developer decides what prompts a student will receive, how their rating is calculated, what a parent sees in an app, what data is collected, and where it goes. None of these choices looks like “policy”; they appear to be technical settings. But taken together, they are policy: they shape how millions of children learn, what counts as success, and what counts as failure. When such decisions are made outside the state contour, school de facto gives part of its sovereignty to whoever wrote the code. And this happens quietly, without a single public decision.

To understand why platforms matter so much, we need to understand what feeds them: data.

Artificial intelligence in education does not work in a vacuum. For AI to build a personal trajectory, diagnose gaps, and predict difficulties, it needs long, continuous, structured datasets — essentially, the digital educational history of a child accumulated over years of study. This is not a one-time file, but a living resource formed throughout the school path and becoming critical infrastructure.

The key implication follows from this. If these data are fragmented, scattered across incompatible systems, or lose continuity, the functionality of AI drops sharply and managerial and social risks rise. An educational history broken into pieces is useless — like a medical record in which half the pages are missing and the other half are written in different languages.

This leads to an important principle: basic educational data and the integrity of a child’s educational history must remain within the state’s sphere of responsibility. This is not a matter of ideology, but of reliability. The state is the only actor that exists long enough and is accountable to society enough to guarantee the preservation of educational history across all years of schooling and to hold it as a public good, not a commercial asset. Here a natural question arises that must be addressed honestly: whose history is it — the child’s, the school’s, the state’s, or the platform’s that collected it? The answer is part of the institutional architecture that is still missing.

Now to the dilemma itself. It has no simple solution, however much one might want one.

On the one hand, responsibility for data and educational history should lie with the state. On the other, private digital platforms have what state systems usually lack: technological flexibility, expertise, and speed of development. The private sector knows how to make products quickly and well. But — and this is the crucial point — its technological potential directly depends on access to data. Without data, even the most talented developer is powerless.

This is the knot: the state bears responsibility for data, while much of the technology and innovation is born in the private sector, which needs those data. How can these roles be combined? How can developers be given the ability to create good products without handing them the educational histories of millions of children as a resource to exploit? This is the institutional dilemma, and the distribution of responsibility remains an open question.

It is tempting to swing to one extreme. Either “the state will build everything itself,” producing slow, inflexible systems that lag behind life. Or “let the market do everything,” turning children’s educational histories into a commercial resource and losing governability. The answer is not at the poles. It lies in building an architecture of interaction in which private platforms can work safely and predictably, while the state retains control over data and responsibility for it. The problem is that such an architecture does not yet exist: public and private solutions are developing in parallel, without agreed data standards, interoperability, and allocation of responsibility. This void must be filled.

Anyone who tries to untie this knot will immediately run into law. The picture here is complicated.

Kazakhstan has just formed a fundamentally new regulatory framework for the digital environment. In November 2025, the Law on Artificial Intelligence was adopted and entered into force on January 18, 2026; Kazakhstan became the second jurisdiction in the world after the European Union to have a separate law on AI. In early January 2026, the Digital Code was signed and entered into force — a large set of rules for the digital world that for the first time systematically enshrined citizens’ digital rights, data regimes, and principles of digital ethics. These are serious and necessary steps. But precisely because the framework is new, it also has a reverse side: the laws are already in force, while established enforcement practice does not yet exist. Many new concepts still have to be tested in real situations, and until that happens, legal uncertainty remains, especially painful for those who work with educational data.

The Law on AI establishes the right principles: transparency and the right of a person to know that an algorithm participated in a decision; explainability; non-discrimination; labeling of AI-generated results; risk management; and responsibility of system owners. The Digital Code adds protection of personal data, digital inclusiveness, and the right to control one’s data. On paper, everything is coherent. Difficulties begin where these norms meet school reality.

One illustrative contradiction shows this best. On the one hand, the system must preserve a student’s educational history — without it, as we have seen, AI does not work, and educational continuity itself requires memory. On the other hand, a person has the right to protection and control of personal data, including deletion. What should happen when a parent demands, “delete all data about my child,” while the system replies, “we are obliged to store it”? There is no established way yet to resolve such conflicts. This is only one example among many.

There are other traps as well. The requirement of explainability is valid and legally enshrined, but it collides with technical reality: modern models can be so complex that explaining every decision is simply impossible. The requirements of non-discrimination and human review sound impeccable, but depend on weak local governability. The most popular school AI scenarios — automatic behavior monitoring and anti-cheating systems — can easily conflict with the protection of sensitive data of minors and restrictions on surveillance. Add the administrative burden — risk management, documentation, audits — for which the system is not yet ready. Correct norms, without mature processes and practice behind them, risk either paralyzing useful projects or remaining declarations that do not prevent spontaneous use.

Here it is worth dispelling a common misconception that often pushes decisions in the wrong direction. Many assume that implementing AI automatically reduces costs: install an algorithm and save on people. In practice, the opposite is true.

Adapting digital systems for different categories of users, ensuring digital inclusiveness so that services can be used by everyone including children with disabilities, security, reliability, and constant support inevitably increase both complexity and cost. AI often does not make systems cheaper; it makes them more expensive, especially when scaled across an entire system. Financial planning for AI transformation should therefore proceed not from mythical savings, but from long-term growth in costs for safety, support, and scale.

This is compounded by the scale of the system itself. Millions of students, hundreds of thousands of teachers — any decision is immediately replicated across everyone and has broad social consequences. In the world of technology startups, an error affects a limited group of users and is easy to fix. In education, the cost of error is social and long-term. School cannot be a testing ground for unlimited experiments; it requires institutional caution and staging.

Interestingly, the market itself understands this better than public slogans suggest. If one looks at what Kazakhstani EdTech is actually doing, rather than what appears in presentations, the picture is soberingly modest.

In practice, mainly limited and pragmatic functions are being implemented: performance analytics, automation of routine operations, basic recommendations, and assistants. No replacement of the teacher by a robot — just ordinary, useful, grounded work. Even these modest solutions require serious computing and financial resources; each pilot project is expensive, which explains the caution of players. Public techno-optimism and actual market practice diverge: revolution in words, careful steps in reality.

This is why participants in our studies propose not to rush, but to create protected testing regimes — technological and managerial “sandboxes” where experimentation is possible without destructive consequences. This is a reasonable response to the dilemma: give innovation space, but fence it off so that an error does not become a social catastrophe.

One more issue cannot be avoided: digital sovereignty. Successful implementation of AI is impossible through simple copying of foreign solutions. The country needs its own models that account for the Kazakh language and context — such as the national language model KazLLM developed by domestic researchers — and its own infrastructure. Legislation has already provided for a national AI platform for developing and testing domestic models and safely storing data; an international AI center is also being developed. Thoughtful partnerships with external actors are needed on Kazakhstan’s own terms. Global technologies can and should be used, but with strict control over data and with reliance on a national foundation. Otherwise, the architecture of our school will be designed for us — and not necessarily in our interests.

Let us bring everything together. The main conclusion of this material is simple in form and difficult in execution.

AI in school rests on data and platforms, which means that the outcome of transformation is decided not by the choice of the “best application,” but by who builds the architecture of interaction between the state, private platforms, and schools — and how. Without such an architecture, the process will continue to reproduce fragmentation, legal uncertainty, and growing risks even with the best intentions and generous funding.

The role of the state in this architecture is special. It does not have to — and will not be able to — develop all solutions, operate every tool, or control every classroom. But it must be the architect of the framework, the guardian of the integrity of educational data, and the guarantor that the rules of the game are fair for the market, the school, and the family. The state does not govern technologies. It governs the balance and holds what cannot be handed over to anyone else.

We have come a long way: from the mandate of the Year of AI through global trends, the voices of our system, other countries’ mistakes, and an honest diagnosis — to this hardest institutional knot. One question remains: what should be done with all this? What might a managed, staged, and human-centered transition look like? This is the subject of the final, seventh material.

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