Generative AI Improves Homework Scores but Lowers Exam Results, Study Finds

A multi-year study of secondary school students indicates that while artificial intelligence tools boost homework grades and cut completion times, they lead to a significant drop in exam performance when students are tested without technological assistance.

The working paper, conducted over 30 months by David Stromberg of Stockholm University alongside Victor Lei and Yanhui Wu of the University of Hong Kong, tracked approximately 27,000 students aged 12 to 18 across nine academic subjects in China. The researchers combined data from daily homework assignments, monthly closed-book examinations, and high-stakes entrance tests to evaluate learning outcomes.

The results revealed a clear divergence between completing schoolwork and retaining knowledge. Following the adoption of generative AI chatbots, students saw their average homework scores rise by 18 percent. At the same time, the average duration spent completing each assignment dropped by roughly 30 percent, decreasing from 64 minutes to 45 minutes.

However, when these students sat for monthly closed-book exams where AI assistance was prohibited, their test scores fell by 20 percent within six months. Over a longer timeframe, researchers observed a similar decline in student performance on high-stakes entrance examinations.

The study’s conclusions align with research published in January by the Organisation for Economic Co-operation and Development. In its Digital Education Outlook 2026 report, the organization noted that offloading tasks to generative AI models without structured educational frameworks improves assignment results without generating genuine learning gains. Research co-authored by academic Dragan Gasevic characterized this loss of productive mental effort as “metacognitive laziness”.

Additional survey data illustrates how youth use AI tools when studying independently outside of school hours. An Ipsos poll conducted for the Vodafone Foundation across seven European countries surveyed 7,000 students aged 12 to 17. It found that 56 percent used AI to gather information and 45 percent sought explanations of terms and concepts. However, 31 percent reported using AI tools to produce complete answers for their assigned tasks. Only 29 percent used interactive features to guide learning, while 20 percent relied on AI to create personalized plans and monitor progress.

Artificial intelligence is also affecting broader student experiences in higher education. A survey of 1,054 full-time UK undergraduates by HEPI and Kortext revealed that 49 percent believed generative AI improved their university experience, citing timesaving and instant support. Regarding social well-being, 59 percent reported AI had no effect on loneliness, while 21 percent felt less lonely and 20 percent felt more isolated.

Background

Governments worldwide are increasingly formalizing artificial intelligence instruction within national educational frameworks. China mandates that primary and secondary students receive at least eight hours of AI instruction per year. Similarly, several Gulf nations—including the United Arab Emirates, Saudi Arabia, and Qatar—have integrated AI concepts into their school curricula as the technology becomes widespread among youth.