AI in Learning Report 2026

AI use in education is now close to universal; its benefit for learning is not. Drawing on about 270 distinct sources, from studies and official reports to surveys and news, this second WAC Annual Report finds that open access to AI often improves the work students hand in while weakening what they learn, whereas tutors designed to withhold answers, used with teachers, can produce large gains. Rigorous evidence is thin, institutions trail their users and assessment bears the costs. The report introduces the WAC AI Learning Readiness Framework and ten signals to watch.

The report in numbers

Key findings

1. Use is near universal: 95 percent of UK undergraduates used AI in HEPI's 2026 survey, up from 66 percent in 2024, and US students using AI for homework rose from 48 to 62 percent in seven months of 2025.
2. In a PNAS field experiment, unguarded GPT-4 raised high-school practice scores by 48 percent but cut later unaided exam scores by 17 percent; a guardrailed tutor removed the harm without adding a gain.
3. In a 2026 randomized trial (working paper) with 2,379 University of Maryland undergraduates, access to a course-integrated AI tutor lowered final grades by 0.37 standard deviations, and only about 15 percent of students used it.
4. In a working paper following 26,811 Chinese secondary students, AI adoption raised homework scores 18 percent but cut monthly exam scores 20 percent, with entrance-exam losses of 18 to 24 percent over about two years.
5. Structured, purpose-built AI tutoring can work: gains of 0.73 to 1.3 standard deviations in a Harvard physics trial and 0.31 standard deviations in six weeks in Nigeria (World Bank working paper), at about 48 dollars per pupil.
6. The evidence base is thin: only 20 of more than 800 K-12 AI papers offered strong causal evidence, and a widely cited positive meta-analysis (g = 0.867) was retracted in April 2026.

The report has 12 key findings in all, 12 chapters of evidence, an original framework and recommendations for five audiences.

Inside the book

1. 
   Chapter 1

   AI in learning in 2026: what changed and how to read the numbers

   AI use is now easy to count and learning is still hard to measure; this chapter defines the field and the traps in its most-cited numbers.

2. 
   Chapter 2

   Adoption in 2026: students, teachers and institutions

   Student use is near-universal and teacher use is climbing, but policies, training and licenses trail behavior, and the gaps by place, gender and income persist.

3. 
   Chapter 3

   Does AI help people learn? The evidence to date

   Purpose-built AI tutors can produce large gains, but in 2026 the evidence turned sharply: open access to chatbots often raises homework scores while lowering real learning.

4. 
   Chapter 4

   Teachers and faculty: time, training and a changing job

   Trials confirm AI saves teachers time, though far less than they report, while training, guidance and safeguards on marking and professional skill trail well behind use.

5. 
   Chapter 5

   AI in education around the world in 2026: a regional analysis

   Governments are racing to put AI in classrooms, but the 2026 record shows reversals, thin evidence and a widening gap between policy ambition and school reality.

6. 
   Chapter 6

   Universities: strategy, licenses and the new campus

   Campuses have bought AI faster than they have planned for it; the real changes are coming in assessment and the entry-level job market.

7. 
   Chapter 7

   Schools: from bans to AI literacy

   Schools moved from blocking ChatGPT to teaching AI in three years, but guidance, teacher support and evidence trail student use, especially in poorer schools.

8. 
   Chapter 8

   Risks and trade-offs: integrity, thinking, equity and wellbeing

   Unstructured AI use can lift grades while cutting learning, detectors misjudge the wrong students, and companion chatbots have become a child-safety issue regulators are only starting to address.

9. 
   Chapter 9

   The AI edtech market: money, products and evidence

   Capital, users and pricing power have shifted to AI labs and platforms; answer businesses collapsed, and evidence of learning impact still trails adoption.

10. 
    Chapter 10

    Governing AI in education: law, guidance and standards

    Binding law, federal promotion and soft guidance now coexist; 2026 brought delays and reversals, and institutions still trail their own users.

11. 
    Chapter 11

    The WAC AI Learning Readiness Framework

    A practical tool, synthesized from this report's evidence, for judging whether an institution or system is ready to use AI in ways that improve learning.

12. 
    Chapter 12

    Outlook 2027: ten signals to watch

    Ten signals will show whether AI is improving learning or only the work students hand in, and what evidence would change that reading.

A chart from the report

One recommendation for each audience

This report is desk research: a synthesis of about 270 published studies, official reports, surveys of students, teachers and institutions, and journal analyses. WAC ran no survey or trial for it. Every chapter was checked against its sources by an independent fact checker before publication. Edorer Education and Lecture Studio, described in Chapter 9, are related to WAC; they are described from their own public pages. The photographs are AI-generated illustrations. Send a correction