Artificial intelligence (AI) is becoming ubiquitous, and classrooms are no exception. While generative AI tools such as ChatGPT can increase productivity and be used to make common tasks more efficient, their effects on how humans learn and approach new tasks could prove catastrophic in the future. The problem is not the use of AI itself, but how it is used. If we take into account how human learning works, it may be possible to reap some of the benefits of AI without undermining future human performance.
A recent study by Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Geb, Özge Kabakcı, and Rei Mariman from the University of Pennsylvania (United States) and the British International School (Hungary) provides some clues about the dangers of using AI in educational contexts and how to reduce them. In this study, published in the scientific journal Proceedings of the National Academy of Sciences (PNAS), the researchers tested how the use of generative AI affected mathematics learning in high school students.
The experiment involved around 1,000 students, divided into three groups. One group of students received an AI tutor similar to ChatGPT-4 (GPT Base); another group received a tutor based on ChatGPT-4, but adapted by researchers to include learning protections (GPT Tutor), such as giving clues to solve problems rather than the final answer; a third group served as a control and used only textbooks and their notes. The researchers developed four 90-minute sessions covering about 15% of the math curriculum for 9th, 10th, and 11th graders. In each session, teachers began by reviewing a previously taught topic; then, students solved exercises according to each of the three conditions described above; after submitting their answers, teachers quickly reviewed the correct answers; finally, students took a no-consultation test with problems similar to those in the study phase.
During the exercise-solving phase, students answered questions such as, “Find the equation of the line passing through A(-2,3), parallel to 2x-3y+5=0.” GPT Base could provide the complete solution to this problem. In contrast, GPT Tutor had been instructed by researchers to encourage students to find the solution themselves and to ask them to begin the problem-solving process before clarifying possible confusion or helping students reach the next step in the solution. GPT Tutor had also been instructed to start by giving as little information as possible and to gradually increase the information provided.
The results indicated that, during the study phase, students who used GPT Base or GPT Tutor increased their performance more than students who only used their notes and textbooks: GPT Base increased performance by about 48%, and GPT Tutor by about 127%.
However, do these effects on performance mean that AI helped students learn? The answer is no, quite the contrary. In the final tests, students who had used GPT Base saw their performance decrease by about 17% compared to students who had not used AI to study. And students who had used GPT Tutor did not differ from those who had not used AI.
The researchers point to two mechanisms that could cause GPT Base to hinder learning: 1) GPT Base generates errors that influence students; and 2) the use of GPT Base prevents students from actively trying to understand problems before attempting to solve them. By analyzing the responses of GPT Base and students, researchers concluded that the second mechanism is the most likely: AI is used as a crutch that prevents students from actively trying to understand problems, thereby hindering learning. The fact that students almost always immediately ask GPT Base for the final answer also supports this conclusion. In the case of GPT Tutor, students could ask for clues but did not receive the final answer, which prevented GPT Tutor from being used as a crutch — hence it did not negatively affect learning, compared to the control group.
Another interesting finding is that, compared to students who did not use AI, students who had used AI overestimated their learning and performance on the final test. This effect of AI on metacognition is a learning illusion that can cause students to stop studying before they have actually learned and prevents them from focusing on the topics they find most difficult, with negative repercussions for learning.
In summary, using a tool similar to ChatGPT during learning impaired learning and students' ability to identify their level of learning. Adding learning protections to this tool eliminated the negative effects on learning. However, not only did it not improve performance, it created the same illusions of learning as ChatGPT.
This study shows that the use of AI in educational contexts should be monitored by the teacher and include learning protections to ensure that incorrect answers are not generated and that students do not use AI to give them final answers without trying to understand the problems. However, in addition to the implementation of these safeguards requiring a lot of work from the teacher, the use of AI did not improve performance and even led to learning illusions. Thus, it seems that the negative effects of AI in the classroom continue to outweigh its potential benefits.
But what can teachers and parents do when the use of AI is almost inevitable? One option is to create activities that cannot be entirely solved by AI and show students the mistakes that AI makes, emphasizing the use of learning strategies that increase learning ability and critical thinking. It is important that educational practices emphasize the importance of “learning to learn” and no longer focus solely on content acquisition.
Some authors have suggested that over-reliance on AI can undermine mechanisms essential to learning, reasoning, critical thinking, and creativity (Oakley et al., 2025). For example, memory is one of these essential mechanisms: we know that the act of memorizing and reproducing what we have memorized (the practice of retrieval) is one of the most effective strategies for learning. However, if AI replaces retrieval practice and other strategies that make our learning more effective, in the long term we may jeopardize our ability to learn, solve problems, detect errors, and transfer learning between contexts.
Even though AI is here to stay and can bring many benefits, it seems certain that these benefits will only be enjoyed by those who are able to assess the limitations of AI and actively monitor and regulate their mental activity to solve problems (i.e., metacognitive ability). An example of this is a study that indicates that the use of AI can increase creativity in work contexts, but only for workers with high levels of metacognitive ability (Sun et al., 2025).
Therefore, education should continue to focus on developing metacognitive skills and using strategies that maximize the ability to learn and solve problems independently. It should also include digital literacy classes that highlight the risks of using AI and the errors that AI generates.