AI Boosts Homework Scores but Hurts Exam Performance, Study Finds

AI's Double-Edged Sword in Education

Artificial intelligence has become an inseparable part of student life. Surveys show that between 80% and 94% of undergraduates in wealthy countries now use AI tools for their studies. But a new large-scale study reveals a troubling paradox: while AI dramatically improves homework performance, it simultaneously undermines actual learning.

The research, tracking nearly 27,000 Chinese middle and high school students over 30 months, found that AI adoption raised homework scores by an average of 18% while slashing completion time by 30%. Yet the same students saw their closed-book exam scores plummet by 20% within six months. The findings suggest that AI may be breaking the traditional link between homework effort and knowledge retention.

The Study: What the Data Shows

Conducted by David Strömberg of Stockholm University and Victor Lei and Wu Yanhui of the University of Hong Kong, the working paper titled "The Generative AI Learning Penalty" analyzed students aged 12 to 18 across nine subjects. Approximately 80% of participants reported using Chinese AI models like Doubao and DeepSeek, with the remaining 20% serving as a control group.

After six months of AI use, homework scores rose by 18%, and average assignment time dropped from 64 minutes to just 45 minutes. However, monthly closed-book exam scores fell by 20% compared to non-AI users. The negative impact intensified over time, with high-stakes entrance exam scores declining by 18% for the Gaokao and 24% for the Zhongkao after roughly two years of sustained AI use.

Why Homework Improves but Exams Suffer

The disconnect likely stems from how students use AI. Those who effectively delegated their homework to AI saw the most severe learning losses. Researchers note that AI can serve as a powerful thinking aid, helping students understand complex concepts and work through problems. But when it becomes an answer-generating tool, it bypasses the cognitive effort required for knowledge consolidation.

This aligns with concerns raised by sci-fi author Ted Chiang, who describes AI as a "cognitive crutch" that hinders intellectual development. A separate study on ChatGPT found that using it for assignments significantly impaired long-term knowledge retention by reducing cognitive effort. The pattern suggests that AI optimizes for task completion rather than learning.

continue reading below...

Nuance: AI Isn't Inherently Harmful

The researchers caution against painting AI as purely detrimental. Students who spent roughly the same amount of time on homework as non-AI users achieved similar exam scores, despite having higher homework grades. This indicates that AI was less harmful when used to assist with work rather than replace it entirely.

The effect also appears to be concentrated among high-achieving students, who may be more likely to rely on AI for shortcuts. This mirrors historical patterns in educational technology, where new tools often amplify existing student differences rather than uniformly improving or harming outcomes.

Implications for Educators and Students

The findings carry significant implications for schools and universities grappling with AI policies. Homework grades may no longer be a reliable indicator of student knowledge if AI tools are permitted. Educators may need to redesign assessments to emphasize in-class, closed-book testing or incorporate AI-use disclosure requirements.

For students, the lesson is clear: AI should augment thinking, not replace it. Using AI to generate ideas, explain concepts, or provide feedback can enhance learning. But outsourcing entire assignments to AI risks short-term gains at the cost of long-term knowledge, especially on high-stakes exams that determine college admissions and career trajectories.

The Bigger Picture: AI and Cognitive Fitness

Beyond education, this study adds to a growing body of evidence that AI reliance may reduce "mental fitness." As AI becomes ubiquitous in workplaces and daily life, the ability to think independently and solve problems without machine assistance becomes increasingly valuable. The study's authors stress that AI is a tool, not a substitute for learning, and its benefits depend entirely on how it's used.

As schools and policymakers navigate this new landscape, the challenge will be harnessing AI's productivity gains without sacrificing genuine understanding. The data from this study offers a clear warning: what gets measured improves, but what gets skipped—deep cognitive engagement—ultimately matters more.