Study finds an 18 percent jump in homework scores and a 20 percent drop in exam results

A recent study tracking 27,000 students in China found that relying on artificial intelligence tools boosts homework grades, but leads to a marked decline in performance during real examinations conducted without technical assistance, according to a report published by The Economist on August 18. The research, led by David Stromberg of Stockholm University alongside Victor Li and Wu Yanhui of the University of Hong Kong, followed students aged 12 to 18 in an environment marked by the rapid adoption of AI applications among adolescents.

Data from the study showed that roughly 80 percent of participating students reported using AI models such as Doubao and DeepSeek, while the remaining 20 percent served as a control group that did not rely on the software. Over a six-month period, average homework scores for students using the technology rose by 18 percent across all academic subjects.However, invigilated tests conducted in exam halls without assistive tools showed the opposite outcome, with those students scoring 20 percent lower than peers who did not rely on artificial intelligence.

These findings arrive amid expanding reliance on technology in education. An earlier survey by the educational technology company Chegg revealed that 80 percent of university students in wealthy nations use AI in their studies, while more recent surveys showed the figure reaching 94 percent in Britain and 93 percent in Germany. The Economist noted that teachers have observed an increase in formulaic, uniform essays suspected of being generated by bots such as ChatGPT, pointing to the scarcity of rigorous evidence regarding the real impact of these tools on learning outcomes prior to this extensive study.

Apparent gains in interim coursework do not automatically translate into conceptual understanding during closed assessments.This observation aligns with a 2024 experiment conducted by the University of Pennsylvania in mathematics classes, where students practiced problem solving using traditional notebooks and textbooks while other groups relied on ChatGPT and intelligent tutoring software. The results showed that technology users outperformed their peers in short-term practice sessions, but that advantage did not extend to subsequent closed-book exams.

The research paper prompted discussions on X and Hacker News, where commentators attributed the testing gap to copying ready-made answers without deep cognitive engagement with the curriculum. Meanwhile, a guidance paper from the Brookings Institution warned that excessive reliance on technology as a replacement for thinking or social interaction limits the development of students' cognitive skills. The full study by the researchers from Stockholm University and the University of Hong Kong had not yet undergone peer review by other academic institutions when the magazine's report was published.