When ChatGPT launched in late 2022, educators quickly wondered whether students would use artificial intelligence to cheat, study, or simply do their homework more efficiently. There is evidence of a troubling answer: many students appear to complete tasks more quickly but learn less from them.
This conclusion comes from one of the largest studies of how generative AI is changing students’ behavior and academic skills. Sina Rismanchian, a graduate student at the University of California, Irvine, worked with researchers at McGraw Hill to analyze millions of student interactions with ALEKS, an online math platform used by more than four million students from fifth grade through college each year. Because ALEKS includes both low-stakes practice exercises and college placement tests, researchers were able to compare student behavior and performance before and after ChatGPT arrived.
To isolate the effects of AI, the researchers compared two types of math problems that differ in how easily students can outsource them to AI: word problems and graphics problems.
Word problems can be copied and pasted directly into AI chatbots for instant answers. Graphics issues are far more cumbersome. A student would have to upload a screenshot and still recreate the chart in ALEKS using its tools.
After the introduction of ChatGPT, student behavior and performance on the two types of problems began to diverge.
Starting in early 2023, students will spend less time on word problems while continuing to spend about the same amount of time on graphics problems. The gap increased from quarter to quarter. By the end of the study period, near the end of 2025, the average time spent on word problems had fallen by 31 percent for high school students and 27 percent for college students – from about four minutes per word problem to less than three. (Middle school students saw only a modest 9 percent decline and fifth graders saw virtually no decline.)
The researchers believe these averages are being dragged down by some students who spend only seconds on word problems because they use AI to answer them.
The same pattern emerged in university placement tests. When exams were taken unproctored, students spent much less time on word problems after ChatGPT was released. On proctored exams, time spent on word problems returned to historical norms.
But time is only half the story. The more disturbing finding is what happened to learning.
Many colleges allow new students to retake placement tests after practicing more math in ALEKS, giving them a chance to qualify for a higher level course. Before ChatGPT, this practice generally paid off. After ChatGPT, students answered more word problems correctly in unsupervised practice sessions, but performed significantly worse on the same problems when they later took a proctored placement test.
In the past, students have answered about 80 percent of these word problems correctly on supervised placement tests. After ChatGPT was introduced, that number dropped to about 60 percent – a reduction of about 25 percent in the probability of solving a word problem correctly.
In contrast, performance did not decrease with graphics issues.
Table of Contents
After ChatGPT was released, students performed worse on word problems (AI-prone) on proctored exams, but solved more word problems correctly in unproctored environments
If students’ math skills had generally deteriorated due to pandemic learning loss, weaker high school preparation, or digital distractions, graphics performance should have deteriorated as well. That wasn’t the case.
The study cannot definitively prove that students used AI. The researchers couldn’t see what else was happening on students’ screens outside of ALEKS. But it is difficult to imagine any other explanation. The changes occurred only for problems that could be easily outsourced to AI, disappeared under supervision, and grew steadily over nearly three years.
“What makes me nervous is that it’s not just about the word problems,” Rismanchian told me. “This cognitive surrender could happen in writing, in science, in anything.”
The paper “Faster Completion, Less Learning” was published as a working paper in June 2026 and has not yet been peer-reviewed. Like every single study, it does not answer the question of how often students use AI in their school work, whether and to what extent it affects learning. But it joins the growing body of evidence that generative AI is causing students to skip the brain work that leads to learning, and that this “cognitive capitulation” is becoming the norm.
A randomized experiment in Turkey found that high school students who used AI to learn math ended up learning less than students who trained without AI. Anthropic, the creator of Claude, has separately reported that many students appear to be using AI to get answers and relieve cognitive work. Rismanchian’s previous study, published in March 2026, documented troubling patterns of AI use in short-answer essays among students at a large California research university.
This does not mean that AI always undermines learning. Carefully designed AI tutors have improved student performance in controlled experiments by asking questions, personalizing instruction, and withholding answers until students solve a problem. But using AI in this way is likely to increase the amount of time students spend on a problem, Rismanchian said. The ALEKS data shows the opposite.
Rismanchian doesn’t think the answer is simply to ban AI. Instead, he argues, students need to value learning enough to resist the temptation to outsource it.
A recent RAND survey suggests that many already recognize the threat to their brains. Students fear AI is weakening their critical thinking skills as more and more of them admit to using it for school work.
The students are not solely to blame. Although many professors have warned their students against using AI to complete their lectures, universities themselves have embraced the technology, often giving students free access to premium chatbots.
“I think we need to convey to students that you should value your learning,” Rismanchian said. “If ChatGPT does it for you, then you haven’t learned.”
Rismanchian understands the temptation.
As an international student, Rismanchian began using ChatGPT to improve the English in his papers. The ideas were still his own. But after a few months, he said, he noticed something disturbing.
“I realized I couldn’t write anymore,” he said. “I lost my writing skills.”
So he stopped using AI to write.
He still uses it for coding.
Contact the staff Author Jill Barshay at 212-678-3595, jillbarshay.35 on Signal or barshay@hechingerreport.org.
This story about the use of AI undermining math skills was produced by The Hechinger Reporta nonprofit, independent news organization covering education. Sign in Proof points and other Hechinger newsletters.
This article first appeared in the Hechinger Report and is republished here under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
https://hechingerreport.org/proof-points-ai-eroding-math-skills/
