Education
Generative AI Boosts Homework Scores but Lowers Exam Results, World Bank Study Finds
Evidence cited by the World Bank shows that unrestricted use of generative AI can improve what students submit as part of classwork while weakening what they actually learn. The World Development Report 2026 includes a section titled ‘Generative AI and the Erosion of Knowledge, Learning, and Human Skills’, warning that tools such as ChatGPT can reduce the mental effort required for reasoning, writing and knowledge building.
The strongest evidence comes from China. A study of students in grades VII–XII found self‑reported AI use rising from almost zero to about 80% between 2022 and 2025. Six months after first use, homework scores improved by 18% even as students spent less time on assignments, while monthly in‑class examination scores fell by 20%. Students who had ever used AI scored, on average, 5–7% lower on senior high‑school and college entrance exams.
The report cautions that the study has not yet been peer‑reviewed and reflects the early phase of AI adoption. Gaurav Nayyar, who led the World Bank team that prepared the report, said the evidence points to both opportunity and risk in education. "If students outsource the learning process to AI, then… that will be a problem," he told TOI. He added that the evidence so far appeared more encouraging when AI was used by teachers rather than directly by students, noting its use in personalising and customising instruction for different learners.
Other research reviewed by the World Bank points in the same direction. Participants using large language models for research displayed lower cognitive effort and weaker reasoning, while studies found mental effort declining even when output quality improved. Students given unrestricted AI access also performed worse later when the technology was withdrawn.
The report warns that if students use AI to complete homework, the "productive struggle" central to learning can disappear, potentially weakening foundational skills in reasoning and writing. However, it does not argue against AI in education. A synthesis of experiments across China, Ghana, Nigeria, Turkiye and the US found that 30 sessions of AI‑powered tutoring produced learning gains equivalent to at least 0.5 years of schooling, particularly when AI complemented teachers and included safeguards against shortcuts. "AI can augment human capabilities," the report said, "but it must not replace the processes through which more advanced skills are acquired."
Frequently asked questions
How did generative AI affect homework scores in China?
Six months after first use, homework scores improved by 18% even as students spent less time on assignments.
What impact did generative AI have on exam scores?
Monthly in‑class examination scores fell by 20%, and students who had ever used AI scored 5–7% lower on senior high‑school and college entrance exams.
What does the World Bank report say about AI use by teachers?
The evidence appeared more encouraging when AI was used by teachers rather than directly by students, noting its use in personalising and customising instruction for different learners.