← All articles

Mikhail Polyakov

AI at school: where it helps students learn, where it gets in the way, and what is now allowed

Level: BeginnerReading time: 8 min

Topics: AI, Education

A school desk with an open squared exercise book full of geometry sketches and a tablet, with glowing dotted lines running from the screen to the sketches

The short answer: whether a neural network helps with learning depends on how it is used. In one large experiment, school students with an ordinary chatbot did better on practice problems, but on a test without it they did worse than those who had no bot. The same model, set up as a tutor that gave hints instead of ready answers, did not show such a drop. At the end of September 2026, Russia's Ministry of Education (Minprosveshcheniya) issued guidelines: younger students do not use neural networks on their own, from grade 5 they may with the teacher's permission, and no one may use them in tests and exams. Below we look at the rules by grade, what research says and how to come to an agreement with your child at home.

What Russia's Ministry of Education recommends

The document is called methodological guidelines for teachers. It is not a law or an order: it sets a framework, while the specific rules are set by the school and the teacher. According to the text of the guidelines, the final decision is made by the teacher, who is also responsible for the consequences.

Grades What a student may do Who decides
1-4 Students do not use AI services on their own. The teacher may use a neural network when preparing a lesson The teacher
5-9 Search for and organise information, get additional explanations of difficult topics, make visual diagrams, test hypotheses Only with the teacher's permission
10-11 Everything above plus project and research work: data analysis, calculations, visual materials for a presentation The teacher

The prohibitions are the same for all grades:

  • AI may not be used in tests or in interim and final assessments;
  • it may not be used where a task checks the student's own knowledge and skills;
  • a text, solution, presentation or project made by a neural network may not be passed off as one's own work.

A teacher may additionally restrict neural networks in tasks that develop the ability to reason and come up with one's own ideas.

The guidelines also say that parents are informed about the use of AI in learning, and that preference is given to services built on Russian models and to services that are part of state education systems.

What research shows

A ready-made answer can get in the way of learning. In an experiment by researchers at the University of Pennsylvania, published in 2025 in the journal PNAS, about a thousand high school students in Turkey took part during maths lessons. One group got GPT-4 as an ordinary chatbot, another got the same model set up as a tutor: it gave hints and did not hand over the solution straight away, and a third group studied without AI. On practice problems, the group with the ordinary bot did 48% better than the group without AI, and the group with the tutor did 127% better. Then access to AI was removed and the students took a test. The group with the ordinary bot did 17% worse on it than the group without AI. A possible explanation: with a bot that gives the solution straight away, it is easy to skip your own work on the problem. The group with the tutor showed no such drop.

A well-designed tutor can help. At Harvard, about two hundred students in a physics course studied some topics in active classes with an instructor and some at home with an AI tutor. The instructors built the tutor on the same teaching methods as the classes: it led the student step by step instead of giving the answer. With it, the students learned more and spent less time (Scientific Reports, 2025). An important limitation: these were university students and a tutor built specifically for the course, not an ordinary chatbot on a school student's phone.

The experiments tested different things. The first compared an ordinary bot, a tutor with hints and classes without AI: on the test, only the group with the ordinary bot did worse than the group without AI, by 17%. The second compared a specially built tutor with a class taught by an instructor, and did not compare it with an ordinary bot. These are findings for specific subjects, participants and systems, not a general law for all learning.

Children already use it. According to a Pew Research survey in autumn 2024, 26% of American teenagers aged 13-17 had used ChatGPT for schoolwork, twice as many as a year earlier. Teenagers themselves distinguish between tasks: 54% consider it acceptable to use it to research a new topic, and 18% to write essays. These are US figures.

How AI helps teachers

The guidelines allow teachers to prepare lesson notes, assignments, tests and presentations with a neural network. The main condition: the teacher checks everything that was created, from the facts to whether it suits the children's age. A neural network can be confidently wrong, so its material does not go into a lesson without a check.

The time savings are noticeable. In a Gallup and Walton Family Foundation survey in spring 2025, American teachers who use AI at least once a week estimated their savings at 5.9 hours a week on average. 32% of the teachers surveyed used AI that often.

What teachers should not do: according to the guidelines, information about students, parents and staff, grades, written assessments of students' character, behaviour or performance, students' work and official documents is not uploaded to AI services. If you want to see how a service handles checking essays, it is easier to write a practice example for this yourself, without texts or data of real students.

Why an "AI detector" is not a verdict

Sometimes a student's work is checked by a program that estimates whether a text was written by a neural network. Such an estimate cannot be trusted as proof. An example: OpenAI, the company behind ChatGPT, shut down its own detector in July 2023 because of its low accuracy. It recognised only 26% of texts written by AI and wrongly flagged human text in 9% of cases.

If a child is accused of copying from a neural network on the basis of such a program alone, it makes sense to calmly ask them to explain their work out loud or show their drafts. This is fairer to both the student and the teacher.

How to come to an agreement with your child at home

The guidelines regulate school, and at home you set the rules. Here are agreements that follow from the rules and experiments above:

  1. A younger school child uses a neural network only together with an adult. This matches the guidelines for grades 1-4. In its 2023 guidance, UNESCO proposed a minimum age of 13 for independent use of such services in the classroom.
  2. First on your own, then the bot. The child first tries to solve the problem on their own and turns to the neural network when stuck.
  3. Ask for a hint, not an answer. For example: "Don't solve it for me, ask me a question that helps me think it through." This does not turn an ordinary chatbot into the tutor from the experiment with Turkish students: the bot may still give a ready-made solution, which is why point 4 and the check below matter.
  4. Check the answer. A neural network can be wrong in a confident tone. A useful exercise (this is my advice, not a point from the guidelines): find the mistake in the bot's answer.
  5. Do not tell the bot personal things. It is better not to send your surname, address, school or photos.
  6. No AI in tests or in work meant to show what you can do yourself. And tell the teacher honestly if a neural network helped with homework when the teacher allowed it.

These agreements are only a starting point: on their own, neither hints nor house rules guarantee that the material has been learned. So check understanding: ask your child to explain in their own words a problem they worked on with a neural network. If they cannot explain it, that is not proof of copying but a reason to look into it: go through the difficult part together, give a similar problem without AI and, if the difficulty keeps coming back, talk to the teacher.

Summary

AI at school is becoming an ordinary tool with clear limits: younger students do not use it on their own, middle-grade students use it with the teacher's permission, older students also use it in projects, and no one uses it in tests. In the Turkish experiment, on the test without AI the group with the ordinary bot did 17% worse than the group without AI, while the group with the tutor that gave hints showed no such drop. The agreements about using AI at home listed above follow the second approach, but you will still have to check yourself whether your child has understood the material.

I work on AI agents and automation. If you would like to see my projects or discuss your own task, take a look at my portfolio.

Sources