Study of 26,000 students reveals pitfalls of relying on generative AI
A 30-month study of 26,811 students in China found that generative AI boosts homework grades but erodes real learning over time.
For you: Copying answers from AI hurts you at exam time. Use chatbots to explain ideas, not to do the thinking for you.

Handing homework to an AI is easy. You type the question and get a finished answer in seconds. Grades go up and you save time. But is there a hidden cost? A large study from China gives a worrying answer.
It followed 26,811 secondary school students for 30 months, two and a half years. Researchers compared students who used generative AI with those who did not. The paper is titled The generative AI learning penalty: Evidence from Chinese secondary education, by Strömberg, Lei and Wu, published as a CEPR discussion paper. Habr ran a translation of a review of the study by Soren Kaplan, originally written for Psychology Today.
At first, everything looks great
In the short term the results were positive:
- homework grades rose by 18%;
- time spent on homework fell by about 30%.
Teachers and parents could easily call that a win. But the researchers kept watching.
Then real knowledge slips
Over time the picture changed:
- scores on exams taken without aids fell by 20% after six months;
- entrance exam results dropped by 18 to 24%;
- the sharpest decline showed up after about two years of AI use;
- around 80% of students who used AI showed signs of offloading their thinking to the tool: strong homework, weak results without AI.
The most surprising finding: the most capable students lost the most. They were better at prompting AI, so they handed it more of the work.
How the learning trap works
The researchers call this effect a learning trap. Several mechanisms drive it.
Cognitive offloading. This means passing mental work to an outside tool. A calculator or a reminder app is useful. But if you hand off a task that is supposed to build a skill, the skill never develops.
Delayed feedback. Grades improve right away, while the loss of knowledge shows up months or even years later, when the problem is already deeply rooted.
The augmentation trap. Researchers at MIT Sloan (Caosun and Aral) describe a similar effect at work: short-term productivity gains slowly undermine the skills that made those gains possible. A 2026 GoTo survey found that 39% of employees feel over-reliance on AI has weakened their professional skills. Among Gen Z the figure is 46%.
How to avoid the trap
In the study, students who used AI but kept putting in serious effort of their own lost almost nothing in learning quality. So the problem is not the tool. It is how you use it. The authors suggest:
- Use AI as a thinking aid, not a replacement for thinking.
- Keep doing real work yourself alongside AI.
- Ask yourself regularly whether you are still building your own abilities or relying more and more on the tool.
- Parents, teachers and managers should look beyond results and watch how a person's ability to think independently is developing.
What this means for readers in Kazakhstan
For school and university students this is an important warning. There is no AI next to you during the UNT or final exams, so high homework grades cannot replace real knowledge.
What you can do now:
- Try to solve each task yourself first. Turn to AI afterward, to check your work.
- Ask for explanations, not finished answers.
- Once a week, test yourself on a topic without any AI.
- Give the chatbot the role of a tutor:
You are my tutor. Topic: [topic]. Do not give me the final answer. First ask me a guiding question and let me try to solve it myself. If I make a mistake, show me where it is, but do not tell me the correct answer right away.
That way AI does not think for you. It teaches you to think.
Photo: Desk-laptop-notebook-pen (23958719819), www.Pixel.la Free Stock Photos — CC0
Source: habr.com




Comments
Sign in to leave a commentNo comments yet. Be the first.