New research gives us a grim portrait of how rampant AI usage is affecting young students’ academic performance and learning.

In a study that tracked nearly 27,000 students in China between the ages of 12 and 18 (grades 7-12), around 80 percent reported using AI models like DeepSeek and ByteDance’s Doubao. The remaining 20 percent, who abstained from using AI, formed the control group.

What the researchers from Stockholm University and the University of Hong Kong found, as reported by The Economist, point to an alarming trend. After six months, the homework scores of students who used AI rose by 18 percent across all subjects, while the time it took them to complete each assignment fell from 64 minutes to 45 minutes on average — a 30 percent reduction.

But were they really learning, or were they using AI as a crutch to breeze through the assignments? This is where the findings get ominous: the AI-using students earned scores on monthly exams that were 20 percent lower than their classmates who didn’t use the tech.

This upends decades of conventional wisdom. It used to be the case that students who got good homework grades tended to do well on their exams, too. Now it’s the opposite: it’s these supposed best performers who end up doing the worst when it’s time to bust out the blue books.

To those on the ground, this likely doesn’t come as a surprise. A Brown University professor, for example, noticed this academic year that his students did suspiciously well on a take-home midterm, with the average score in the high 90s, and half of the students getting perfect grades. 

Suspecting that AI had a hand in the phenomenon, he confronted his class about their suspected cheating and changed it so the final exam would be taken in-person. His hunch turned out to almost certainly be correct. The average score on the final collapsed to just 48 percent, when previously the average had never dipped below 65 percent. Dozens of students didn’t even bother to take the final — the majority of them being those who had gotten perfect scores on the midterm.

Other research has explored AI’s cognitive effects on learning. An MIT study published earlier this year found that students who used AI to help write an essay showed lower brain activity compared to the students who didn’t. Alarmingly, the AI group struggled to quote their own essays, and when they were later asked to write a paper without AI help, their brain activity stayed at a reduced level. Mirroring those findings, other studies have showed that AI impairs critical thinking and have associated it with memory loss.

This latest study illustrates academic trends rather than cognitive ones, but it dovetails with what those studies are showing. In addition to the plunging exam scores, it also found that the worst learning losses came in social science subjects, followed by STEM and language courses. Intriguingly, the AI users who took around the same amount of time to complete their homework as non AI users had smaller learning losses, though the losses were still present.

The study also adds to the concerns around AI’s role in grade inflation, which has skyrocketed in recent years. Recent research found that that the percentage of A grades in college courses that were considered vulnerable to AI cheating — which were typically in the humanities and engineering — have surged by 30 percent since the release of ChatGPT.

More on AI: Teachers Warn That Students Are Losing the Ability to Think as They Lean on AI for Everything