Blab AI
China Study Finds Generative AI Use Cuts Exam Scores by 20% While Boosting Homework Grades
In a startling new report, Chinese students who turned to generative AI tools for homework found their exam scores slipping.
The study, published on June 18 2026 by Gabriel Demombynes of the World Bank’s Human Capital Project, followed 26,811 pupils in grades 7–12 across a single county for 30 months beginning in 2022. By exploiting the staggered timing of AI adoption, the authors compared outcomes for students who began using AI with those who did not.
Three key effects emerged. First, AI users cut the time they spent on homework by 30 percent. Second, their homework grades rose by 18 percent. Third, after five months of AI use, monthly exam results dropped by 20 percent, and scores on two high‑stakes entrance exams fell by 18 percent and 24 percent. The largest declines appeared in social‑science subjects, followed by STEM and then languages.
In terms of effect size, the 20 percent drop in monthly exam scores translates to a 1.4‑standard‑deviation loss—an unusually large impact in education research. For comparison, a 2025 study of AI‑based tutoring in Nigeria found a learning gain of only 0.31 standard deviations.
The authors acknowledge that other shocks could influence both AI adoption and academic performance. In parts of China, many children live apart from parents who have migrated for work; the World Bank’s recent human‑capital report notes that such separation can reduce achievement. The study’s authors argue that this factor cannot explain the entire pattern, because students who do not use AI rarely experience a 20 percent decline in test scores. Nonetheless, they cannot rule out a minor role for such shocks and suggest that the true effect of AI may be somewhat smaller.
The findings arrive as generative AI tools become increasingly ubiquitous. Smartphone ownership is nearly universal in most countries, and many teenagers already use AI through search engines and chatbots. The World Bank is actively funding pilots that evaluate AI applications in education and health, and the author’s team has produced a Generative AI Evaluation Playbook.
Beyond the classroom, the authors point out that people are using AI for health information, skills development, career advice, and emotional support—often outside the structured environments that researchers can evaluate. They argue that the largest human‑capital effects of AI may come from unregulated, individual use rather than from carefully deployed tools.
The study is likely to serve as a wake‑up call for parents and educators. It highlights the risk that students may rely on AI as a shortcut, improving short‑term homework performance while eroding deeper learning. The authors call for additional studies in other contexts to determine whether the observed pattern is widespread.
In the meantime, the World Bank and other development partners are continuing to monitor AI’s impact on education and health services. The Generative AI Evaluation Playbook is intended to guide future pilots and help institutions assess both the benefits and risks of AI tools.
The study’s results underscore the need for careful oversight of AI use in education and beyond. While AI can offer powerful support, the evidence suggests that unregulated use can undermine learning outcomes. Future research will need to disentangle the effects of AI from other socioeconomic shocks and to identify best practices for integrating AI into learning environments without compromising deep understanding.
The study is still under peer review, and the authors plan to refine their analysis to address potential confounders. As AI continues to spread, policymakers, educators, and researchers must balance the promise of generative AI with the evidence of its potential to erode learning.
The study, published on June 18 2026 by Gabriel Demombynes of the World Bank’s Human Capital Project, followed 26,811 pupils in grades 7–12 across a single county for 30 months beginning in 2022. By exploiting the staggered timing of AI adoption, the authors compared outcomes for students who began using AI with those who did not.
Three key effects emerged. First, AI users cut the time they spent on homework by 30 percent. Second, their homework grades rose by 18 percent. Third, after five months of AI use, monthly exam results dropped by 20 percent, and scores on two high‑stakes entrance exams fell by 18 percent and 24 percent. The largest declines appeared in social‑science subjects, followed by STEM and then languages.
In terms of effect size, the 20 percent drop in monthly exam scores translates to a 1.4‑standard‑deviation loss—an unusually large impact in education research. For comparison, a 2025 study of AI‑based tutoring in Nigeria found a learning gain of only 0.31 standard deviations.
The authors acknowledge that other shocks could influence both AI adoption and academic performance. In parts of China, many children live apart from parents who have migrated for work; the World Bank’s recent human‑capital report notes that such separation can reduce achievement. The study’s authors argue that this factor cannot explain the entire pattern, because students who do not use AI rarely experience a 20 percent decline in test scores. Nonetheless, they cannot rule out a minor role for such shocks and suggest that the true effect of AI may be somewhat smaller.
The findings arrive as generative AI tools become increasingly ubiquitous. Smartphone ownership is nearly universal in most countries, and many teenagers already use AI through search engines and chatbots. The World Bank is actively funding pilots that evaluate AI applications in education and health, and the author’s team has produced a Generative AI Evaluation Playbook.
Beyond the classroom, the authors point out that people are using AI for health information, skills development, career advice, and emotional support—often outside the structured environments that researchers can evaluate. They argue that the largest human‑capital effects of AI may come from unregulated, individual use rather than from carefully deployed tools.
The study is likely to serve as a wake‑up call for parents and educators. It highlights the risk that students may rely on AI as a shortcut, improving short‑term homework performance while eroding deeper learning. The authors call for additional studies in other contexts to determine whether the observed pattern is widespread.
In the meantime, the World Bank and other development partners are continuing to monitor AI’s impact on education and health services. The Generative AI Evaluation Playbook is intended to guide future pilots and help institutions assess both the benefits and risks of AI tools.
The study’s results underscore the need for careful oversight of AI use in education and beyond. While AI can offer powerful support, the evidence suggests that unregulated use can undermine learning outcomes. Future research will need to disentangle the effects of AI from other socioeconomic shocks and to identify best practices for integrating AI into learning environments without compromising deep understanding.
The study is still under peer review, and the authors plan to refine their analysis to address potential confounders. As AI continues to spread, policymakers, educators, and researchers must balance the promise of generative AI with the evidence of its potential to erode learning.
