Recent research published in Proceedings of the National Academy of Sciences It suggests that giving high school students unlimited access to artificial intelligence for math practice may impair their ability to learn. AI chatbots can help students solve problems correctly when the technology is available, but when the AI is removed, these students perform worse on independent exams than their peers who did not use the tool at all.
Generative artificial intelligence refers to computer programs that can instantly create original content, such as conversational text or mathematical solutions, by analyzing patterns in large amounts of existing data. Although this technology has become very popular, its impact on student development remains a subject of debate.
“Generative AI emerged with great excitement about its productivity potential. However, if AI undermines skill development, it could harm both productivity and cognitive performance in the long run. At the time, there was surprisingly little causal evidence for the impact of AI in education,” said study author Alp Sung, assistant professor at the Wharton School at the University of Pennsylvania.
To investigate whether these systems help or hinder human learning, scientists investigated a concept that educators and psychiatrists call cognitive debt or cognitive atrophy. This theory suggests that outsourcing reasoning to machines can impair the brain’s critical thinking abilities. If the computer completes the task, the user misses out on the productive effort required to independently solve complex problems.
This loss of independent problem-solving ability is especially dangerous because AI models frequently generate incorrect information. A person must maintain the necessary expertise to evaluate the machine’s output. If students do not practice these brain teasers and rely on algorithms early in their education, they may not develop the foundational skills needed for future success.
To test how different software configurations affected these results, scientists conducted a randomized controlled trial at a large Turkish high school during the fall 2023 semester. The sample included nearly 1,000 students in approximately 50 regular classrooms in grades 9, 10, and 11. The researchers focused on mathematics instruction, devoting four 90-minute sessions to topics that accounted for about 15 percent of the semester’s curriculum.
The students were divided into three groups. The first group served as a control and completed the practice questions using only the standard textbook and class notes. The second group was given a laptop loaded with a program called GPT Base. This program works like a standard ChatGPT interface and was instructed to simply act as a math tutor.
The third group was given a laptop with a program called GPT Tutor. This application used the same interface but included invisible background instructions designed by the teacher to protect the learning process. These specific instructions led the AI to provide step-by-step hints rather than direct answers, giving the AI the correct solution to avoid making up false information.
Each class session was divided into three different parts. First, the teacher gave a traditional lecture and we solved example problems written on the board. Next, students engaged in a supported practice period in which they worked on specific problems designed to reinforce lecture concepts.
In this second phase, students used the resources assigned to their specific group, whether it was a textbook or one of two AI programs. Finally, students completed an unaided exam that included questions that were conceptually identical to the exercises. During this final stage, all students worked completely independently without access to laptops, books, or notebooks.
To ensure students are taking the assignments seriously, both the practice exam and final exam grades count towards the actual class grade. To eliminate potential teacher bias, independent graders graded the submitted papers using a standardized rubric.
Researchers observed a significant improvement in performance during the practice phase for students using artificial intelligence. Compared to the control group, students who used the unrestricted GPT base scored 48% higher and students who used guided GPT tutors scored 127% higher.
When researchers evaluated the independent trial results, a very different pattern emerged. Students in the unrestricted GPT Base group performed 17% worse on the unaided exam than students in the control group, who had no access to a computer. Students in the guided GPT instructor group performed statistically the same as the control group, indicating that software constraints prevented the learning loss seen in the unrestricted group.
“Honestly, I was surprised by the main results! I did not expect that AI would have a negative impact on student learning (we don’t usually conduct studies expecting a treatment to have a negative impact on the results!),” Sungu told PsyPost.
“The effects are nontrivial and go in opposite directions depending on what you measure. With access to AI, students scored 48% higher on practice exams, but when that access was removed, the same students scored 17% lower on exams than students who never used AI. These are not subtle differences; This suggests that how students interact with students fundamentally shapes what they remember.
By analyzing chat logs between students and artificial intelligence, scientists have determined why unrestricted artificial intelligence has a negative impact on learning. The majority of students using GPT Base simply asked the program for the correct answer. Although the AI did its duty, the researchers found that the AI only provided the correct answer 51% of the time and frequently made logical and arithmetic mistakes.
Students were blindly copying these flawed answers onto practice sheets without fully understanding the content. They were using the tools as a crutch and were completely unprepared when they had to solve similar problems themselves. In contrast, students using GPT tutors were forced to interact with the material, ask for help, and try to solve problems on their own, which maintained their learning ability.
“In traditional education settings, students tend to use standard generative AI tools as answering machines, a way to get homework done rather than learn,” Sungu says. “That shortcut can have a negative impact on actual learning.”
The researchers also found a significant discrepancy between how much students thought they had learned and their actual performance. Students who took advantage of the unlimited GPT Base program felt they learned just as much as their peers, even though they had a high rate of failing independent exams. On the other hand, students using a guided GPT tutor perceived a significant improvement in their performance, even though their final exam scores matched only the control group.
A common misconception about these findings is the assumption that well-designed artificial intelligence cannot improve educational outcomes at all. Although the study’s guided AI program did not allow students to outperform a control group on the final exam, it used relatively simple background programming. Researchers suggest that more sophisticated systems that act as active tutors could ultimately have a positive effect on learning.
“We tested one tool using very simple prompt engineering,” Sungu said. “Much more work needs to be done on long-term learning, skill development and downstream productivity impacts.”
The main limitation of this study is its specific context, as it only evaluated mathematics instruction in a single high school at a time when AI chatbots were still a novel technology. Scientists note that different subjects, such as writing, lack objective evaluation criteria and may interact differently with artificial intelligence. Furthermore, the researchers only measured short-term outcomes rather than assessing skills over a long period of time.
Future research should investigate the long-term effects of AI use on skill development and see if extended dependence leads to widespread cognitive atrophy. Scientists want to evaluate new educational policies that direct these digital tools to social good rather than intellectual harm.
“AI doesn’t just change the way we work; it changes the way we think,” Sungu explains. “As such, education will be at the forefront. My goal is to study how AI and other digital technologies reshape human capital development more broadly, and to design and test policies that steer these tools toward social value rather than potential cognitive atrophy.”
The study, “Generative AI without Guardrails Can Harm Learning: Evidence from High School Mathematics,” was authored by Hamsa Bastani, Osbert Bastani, Alp Sung, Haosen Ge, Ozge Kabakçu, and Lei Mariman.
