Education and Artificial Intelligence: Universal Student Use, Mixed Evidence on Learning, and the Design Choices That Decide the Outcome

Abstract

Purpose. No industry has seen generative AI reach its users as quickly as education, because students were among the first to adopt it. This article asks where AI stands in education in 2026, what the evidence says about its effect on learning, and where education may be by about 2035. Design and methods. A review combining the share of the education research literature that involves AI (OpenAlex, 2010–2025), national surveys of student use in the United Kingdom, randomised trials of AI tutors in schools and universities in several countries, a meta-analysis that was later retracted and a reanalysis of it, and the EU AI Act as it applies to education. Claims taken from the literature were checked against the sources cited. The scenarios are judgements built on this evidence. Findings. The share of education works that mention AI rose from 0.7% in 2018 to 8.0% in 2025, 12.1 times as much, the largest increase of the 20 industries, and language-model works from 356 to 23,225, 30.8% of AI-related works. Student use is almost universal: in the UK, 66% of undergraduates used AI tools in 2024, 92% in 2025 and 95% in the survey of December 2025, when 94% used generative AI for assessed work and only 36% felt encouraged by their institution to do so. The evidence on learning is mixed and depends on design. In a field experiment with nearly a thousand high-school students, access to a standard GPT-4 interface improved grades while it was available (by 48%) but lowered them by 17% once it was withdrawn, whereas a tutor designed with safeguards raised grades by 127% during practice and largely removed the harm. A randomised trial in a university physics course found that students learned more, in less time, from a carefully designed AI tutor than from an active-learning class. A widely read meta-analysis reporting large benefits of ChatGPT was retracted in 2026, and a reanalysis concluded that the effects diminish once publication bias is considered. We expect AI use to be normal among students and teachers by 2028, assessment and pedagogy to be redesigned in 2029–2032, and the difference between schooling systems to depend on guidance and design more than on access. Originality. The article places the fastest-growing research field and the most rapid user adoption beside a body of evidence that is thin and sometimes unreliable, and identifies tool design, not access, as the factor that the best trials point to.