The AI Learning Paradox
- Editorial Staff
- 5 days ago
- 3 min read
Using generative AI for schoolwork can boost homework grades and save time, but significantly impairs students' actual learning and long-term exam performance. Is there an optimal amount of AI exposure?
Takeaway: This study investigates how students' real-world, self-directed use of generative AI affects both their day-to-day homework performance and their performance on short- and long-term exams.
Key Points
Productivity vs. Learning: Generative AI boosts students' homework scores by 18% and cuts completion time by nearly a third, but causes monthly closed-book exam scores to fall by 20% within six months.
Compounding Long-Term Costs: On high-stakes high school and college entrance exams, scores decline by 18% to 24%, with the full penalty taking up to two years to fully materialize.
Outsourcing vs. Tutoring: Learning losses are concentrated among the 80% of students who use AI to "outsource" work and bypass effort; students who use AI while spending the same amount of time studying experience virtually no learning loss.
Can artificial intelligence make us smarter, or is it making us lazy?
As chatbots like ChatGPT, DeepSeek and others become everyday study tools, parents, teachers, and policymakers face a central dilemma. Is generative AI an on-demand personal tutor that accelerates understanding, or a shortcut that prevents students from developing essential skills?
To answer this, researchers studied how teenagers actually use generative AI when left to their own devices. Unlike short-term laboratory experiments, this study followed 26,811 middle and high school students in China over 30 months as AI adoption surged from near zero to 80 percent.
The researchers tracked digital homework submissions, timestamped completion times, monthly closed-book classroom exams, and high-stakes provincial entrance exams (the Zhongkao for high school and Gaokao for college). By comparing students who adopted AI at different times against those who did not, the authors isolated the true causal impact of AI on learning over time.
Findings
The findings reveal a stark trade-off between task productivity and genuine learning. In the short run, AI acts as a supercharger for homework: within five months of adoption, students' homework scores rose by 18%, while the time spent on assignments dropped from 64 minutes to 45 minutes.
However, this homework efficiency came at a heavy cost. Within six months of using AI, students' scores on monthly, closed-book classroom exams dropped by 20%. On high-stakes college and high-school entrance exams, the erosion was equally severe—falling by 18% and 24%, respectively.
Crucially, while classroom exam grades dropped quickly, the damage to entrance exam scores took up to two years to fully emerge because those exams test cumulative knowledge built over several years.
The Homework Outsourcing Trap
Why does AI harm test scores? The study revealed that roughly 80% of AI users engage in "homework outsourcing"—using AI to generate answers quickly while bypassing the mental effort required to absorb the material. These students completed assignments far faster than even the quickest non-AI users, earning top homework marks but failing closed-book tests.
By contrast, the small minority of students who used AI but continued spending as much time on their homework as non-users suffered no learning penalty at all. This suggests that AI itself does not inherently destroy learning efficiency; rather, using AI as a shortcut removes the cognitive struggle necessary to retain knowledge.
The "learning penalty" appeared across all subjects, but was largest in social sciences (-27%), followed by STEM fields (-22%), and foreign languages (-17%). Younger students, high-achieving students, and boys experienced larger score drops—the latter largely because boys reported using AI more hours per week.
Why It Matters
For parents and educators, the key takeaway is that high homework grades can be deceiving. AI breaks the historical link between homework performance and real understanding. Because AI outputs look polished, teachers and parents may assume a student is mastering the material until a closed-book exam proves otherwise.
To prevent AI from undermining human capital, educational strategies must shift from evaluating final homework outputs to monitoring learning inputs—such as study time, cognitive effort, and closed-book practice.
Learn More
Title: The Generative AI Learning Penalty: Evidence from Chinese Secondary Education
Authors: David Strömberg, Victor Lei, and Yanhui Wu
Publication: Working Paper (2026)
URL: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6868618
Econ Today Explains
Economic Concept: Dynamic Complementarity in Skill Formation
Dynamic complementarity is the principle that "skills beget skills." In education, learning acquired today makes acquiring new skills tomorrow easier and more productive. Learning is a cumulative process where early concepts serve as building blocks for complex ideas.
When students use AI to bypass the cognitive effort needed to master basic material, they create a gap in their foundational knowledge.
Because skills build on one another over time, missing these early foundations makes future learning progressively harder. This explains why the negative impact of AI on broad, cumulative entrance exams takes years to fully materialize—the missed learning compounds over time.
This article was written by AI but reviewed by a real human.
Comments