Two out of three young people in Turkey now use generative AI. In October 2026 the national statistics office reported that 64.9% of 16- to 24-year-olds use tools such as ChatGPT or Gemini, and that use across the whole population nearly doubled within a year, from 19.2% to 37.6%. Nobody has measured how much of that happens at the homework desk. Any teacher will tell you.
The rate is not the real question. Does studying with AI actually teach you anything, or does it just get the homework done? The answer is neither yes nor no; it depends on how you use it. One of the clearest experiments on this was run in a Turkish high school.
The high-school experiment: brilliant in practice, weaker in the exam
During the 2023-2024 school year, researchers from the University of Pennsylvania ran a mathematics experiment with close to a thousand students at a high school in Turkey. Around fifty classes from grades 9 to 11 took part in four 90-minute practice sessions. The classes were split into three groups. The first worked with textbooks and notes only. The second used a chatbot similar to ChatGPT that answered whatever it was asked. The third group's chatbot was the same model, but tuned with input from teachers: it gave hints instead of handing over the solution.
During the practice sessions the AI groups looked great. Students with the unrestricted chatbot did 48% better than the textbook group, and those with the hint-giving version did 127% better. Then everyone sat the same exam, this time with pen and paper and no screen.
The unrestricted group scored 17% lower than the students who had never used AI at all. The hint group scored about the same as the textbook group. When the researchers looked at the chat logs, the reason was plain: the unrestricted group had mostly pasted in the question, asked for the solution, copied it and moved on. The performance that shone in practice was not there in the exam. The preprint, published in 2024, calls this using the chatbot as a crutch.
It is one study at one school, so it should not be read as a law of nature. Still, the result matches what teachers have been saying for two years: the homework is flawless, the exams are not.
Why this happens: convenience removes the fuel that learning runs on
Memory research has repeated one finding for decades: information you produce yourself leaves a deeper trace than information you read ready-made. Struggling with a problem for a few minutes, even if you end up getting it wrong, makes the correct solution stick better once you see it. The mind learns where it works. The mechanism, and how to turn it into a revision schedule, is covered in the guide to spaced repetition and active recall.
A chatbot removes exactly that effort. You paste the question, the solution arrives, you understand it. Understanding feels good. The trouble is that understanding and being able to do something are not the same thing. It is the same student who follows every word in class and then freezes at the first problem at home, only with a different explainer.
A small study shared as a preprint by a team at the MIT Media Lab in 2025 looked at what this does in the brain. Fifty-four participants wrote essays: one group with no tools, one with a search engine, one with ChatGPT. EEG recordings showed the weakest brain connectivity in the ChatGPT group. More striking, that group struggled to quote from the essay it had written minutes earlier and felt less ownership of the text. The sample is small and the work has not been peer reviewed, so it is not a settled result. But it describes something that happens at homework desks all the time: the assignment got handed in, and the knowledge passed straight through.
The wider cost of handing thinking to a tool, from satnavs to calculators, is explored in the article on cognitive offloading. The question here is narrower: while studying, which jobs can you hand over, and which never?
The difference between getting an answer and getting a hint
That was the only difference between the two chatbots in the experiment, and it decided the outcome. You can create the same difference in any chatbot yourself. One sentence is enough: "Don't solve this for me. Give me a hint for the step I'm stuck on."
In 2025 the major chatbots all added an option that works this way: a study mode in ChatGPT, guided learning in Gemini, a learning mode in Claude. These modes ask questions instead of giving answers and move step by step. They help, but they are not enough on their own. They switch off in two clicks, and the moment you want to switch them off is usually the moment you need them most. The discipline is not in the mode. It is in you.
| Finishes the homework | Teaches you |
|---|---|
| Pasting the question and asking for the solution | Trying it yourself first, then saying "no solution, just a hint" at the step where you are stuck |
| Asking for a summary of the topic | Writing your own summary and asking "what is missing or wrong?" |
| Having it write the essay | Giving it your draft and asking "which sentences are unclear?" |
| Reading the solution to a question you got wrong | Pasting your own working and asking "at which step did I go wrong?" |
| Using it as an answer key | Saying "ask me five questions on this topic and grade my answers" |
The right-hand column is slower. That is the whole point. Speed is useful in the exam, not while studying, and the speed you have in the exam is built from the slowness you accept while studying.
Five ways of studying with AI that actually work
1. Having a hard point explained differently. You listened in class, read the chapter, and one part still will not settle. "Explain this with an everyday example, as if to a thirteen-year-old" is exactly right here. The thing to watch is the order: book and lesson first, chatbot second. Do it the other way round and you will never open the book.
2. Generating questions. "Write five questions on exponents, hide the answers, and grade me as I answer." The nice thing about this use is that it moves the chatbot from the answering side to the asking side. Keep in mind that its questions are occasionally wrong; check the arithmetic yourself before trusting its answer key.
3. Diagnosing mistakes. When you get a question wrong, reading the correct solution is the easy road. Writing out your own working and saying "don't tell me the answer, point to the faulty step" is the hard road. The hard road is the only one that stops you making the same mistake again.
4. Making it the student. "I'm going to explain this topic to you. Ask questions like a student who doesn't get it." The gaps in your explanation are the gaps you could not notice while listening. If you can explain a topic, you know it. Where you stumble while explaining is where you need to study more.
5. Handing over the logistics. Calculating revision dates, turning a weekly plan into a table, putting a topic list in priority order. These are secretarial jobs, not thinking. There is no harm in delegating them, and if the time you save goes into solving problems, you come out ahead.
AI makes mistakes too; the checking is yours
Chatbots speak in the same confident tone whether they are sure or not. They can cite a source that does not exist, get a simple multiplication wrong or misremember a date. That is no reason to abandon the tool, but it does fix the roles: you learn the method from it, and you verify the result yourself. Redo the arithmetic in maths; compare factual answers with the textbook. It is a habit that pays off in the exam as well. A student who trusts their own working rather than an answer key hesitates less between the options.
On exam day the AI will not be there
The real lesson of the high-school experiment is not about a tool but about a condition. The exam condition is fixed: paper, pen, a time limit and nothing but your own mind. At least part of your studying has to happen under that same condition, or you will meet it for the first time on exam day. This is why timed topic tests and mock exams have not lost value in the age of AI; they have gained it.
Open the chatbot after the test, not during it, to ask why your wrong answers were wrong. What a test can and cannot do for you, and how to learn from the wrong options, is covered in the benefits of practice tests; how to read a full mock exam is in the practice-test analysis guide.
A note for parents: a ban usually turns into hidden use. Three rules are more realistic: try it yourself first, ask for a hint rather than the answer, and at the end close the screen and test yourself. Most tools now offer parental controls; still, what decides the outcome is not a setting but those three rules.
Frequently asked questions
Is studying with AI harmful?
Not in itself. What harms you is taking the solution ready-made without trying first. In the Turkish high-school experiment, students who did that scored lower in the exam than students who never used AI, while those who studied with hints lost nothing. The difference is in the use, not the tool.
Which AI is best for studying?
All the major tools now have a study mode that asks questions and delays the answer, and their free versions are enough for this. Which one you pick matters far less than whether you make a habit of asking for a hint instead of the answer.
Does having AI do my homework count as cheating?
Your school's rule is the one that applies; some teachers openly allow it, others forbid it. From a learning point of view the answer is clearer: homework is a measurement of what you can do. Homework done by a chatbot gives your teacher false information, and gives you the most.
All of this comes down to one sentence: do not let the AI do anything that will not be there with you in the exam. Do the thinking yourself; take the explanation, the questions and the feedback from it. Keep that order and the tool in your hands is a patient assistant teacher. Break it, and it is the most expensive cheat sheet ever made.