AI in Education: Design Decides Whether Students Learn or Lose

Generative AI can make students sharper, or it can quietly hollow out what they know. A cluster of recent research suggests the outcome depends less on the technology than on how it is deployed. Several randomized trials report strong learning gains from carefully structured AI tutoring. A major World Bank report and a widely discussed MIT study warn of the opposite effect when students use chatbots freely. Read together, they make a single case: the scaffolding around the tool decides the result.

The warning: better homework, worse exams

The sharpest cautionary evidence appears in the World Bank's World Development Report 2026. It includes a dedicated section titled "Generative AI and the Erosion of Knowledge, Learning, and Human Skills." 4 The report argues that tools like ChatGPT can cut out the mental effort through which reasoning, writing and knowledge are built. 4

Its strongest data comes from a study of Chinese students in Grades 7 to 12. Self-reported AI use among them rose from nearly zero to about 80% between 2022 and 2025. 4 Six months after students first used AI, their homework scores were up 18% while time spent on assignments fell. Over the same period, monthly in-class exam scores dropped 20%. 4 Students who had ever used AI also scored 5–7% lower, on average, on senior high-school and college entrance exams. 4

The report itself attaches caveats. The study has not been peer-reviewed, and it reflects an early phase of adoption. 4 Gaurav Nayyar, who led the report team, described the evidence as pointing to both opportunity and risk. 4

The MIT Media Lab's "Your Brain on ChatGPT" study adds a laboratory angle. Researchers split 54 participants into three essay-writing groups: one using an LLM, one using a search engine, and one using no tools at all. 1 They measured cognitive load with EEG and evaluated the essays with NLP analysis, human teachers and an AI judge. 1 In a fourth session, 18 participants swapped conditions. LLM users wrote unaided, and the unaided writers got the LLM. 1 The study's framing around "accumulation of cognitive debt" signals its central concern: offloading effort to an assistant may carry a cost that builds up over time. 1 Its sample is small, especially in the crossover session, so it is better read as a provocative signal than a settled verdict.

The counterpoint: structured AI tutoring works

Other trials point in a very different direction. A randomized controlled trial published in Scientific Reports by Greg Kestin, Kelly Miller and colleagues found that AI tutoring outperformed in-class active learning. 2 Active learning is already a demanding benchmark. Students in the trial also reported feeling more engaged and motivated. 2 The authors stress that their tutor was built on a "novel research-based design." It was not an off-the-shelf chatbot. 2

A trial in Edo State, Nigeria, offers similar evidence from a very different setting. Across nine public secondary schools, students took part in a six-week after-school program. They worked in pairs with GPT-4 under teacher supervision, using prompts written to encourage reasoning rather than shortcuts. 3 Participants posted large gains in English and did better on end-of-year exams. 3 The researchers attribute the success to the surrounding structure: pedagogical guardrails, active facilitation by teachers, and monitoring to prevent over-reliance. 3

Where the evidence converges

At first glance these findings contradict each other. On closer reading they mostly fit together. The negative results describe largely unrestricted use. Students hand assignments to a model, and the visible output improves while the underlying skill weakens. 41 The positive results describe AI built into designed learning routines. The tool prompts thinking instead of supplying answers, and humans stay involved. 23

The Chinese data also shows how misleading the obvious metric can be. If schools judged AI by homework quality, they would conclude it works, even as exam performance moved the other way. 4 The Nigerian and Scientific Reports trials measured learning independently, through end-of-year exams or controlled comparisons. 23 That is the kind of evaluation needed to tell genuine progress from polished output.

There are gaps worth noting. The World Bank section cites no India-specific study of student outcomes, even though its warning is being read as relevant there. 4 The most alarming figures are not peer-reviewed. 4 The positive trials, meanwhile, were bounded programs with researchers closely involved, and scaling them faithfully is a separate challenge.

The takeaway

The defensible conclusion is not that AI helps learning or that it harms it. Generative AI amplifies whatever learning design it is placed into. Left unstructured, it tends to let students skip the effort that builds skill. Wrapped in guardrails, reasoning-focused prompts and teacher oversight, it can deliver tutoring-like benefits at scale. 3 For schools and education ministries, the policy question is not whether to allow AI. It is whether they are willing to invest in the design, supervision and honest measurement that make it work.