AI in Education: the 2025–2026 evidence and forecasts to 2036

Students with access to general-purpose GenAI tools produce better work — but the advantage disappears, and sometimes reverses, on exams once the AI is taken away. Gains in task performance are not gains in learning ✓ OECD 2026.

The OECD names the mechanism: offloading thinking to chatbots breeds "metacognitive laziness" and disengagement from the learning process ✓. Tools built with deliberate pedagogical intent, by contrast, show sustained learning gains ✓.

At the other pole sits the expert-designed tutor: in the Harvard RCT (N=194, introductory physics) students learned more than twice as much from an AI tutor as from an in-class active-learning session, with an effect size of 0.73–1.3 SD ✓. A meta-analysis of 35 experiments (4,193 participants) puts the overall effect at g = 0.670 — and the effect grows with the length of the intervention, so this is not a novelty bump ✓.

• Three quarters of students aged 16+ across the OECD used generative AI in 2025 — the highest of any population group; among all individuals, more than a third ✓.
• Homework with AI (US, RAND): 48% → 62% in just seven months of 2025. Meanwhile 67% of students themselves say AI use may harm their critical thinking — and keep using it ✓.
• Adoption gaps across the OECD: 53.6 points by age, ~21 points each by education and by income, only 4.2 points by gender ✓.
• Teachers: 37% of lower-secondary teachers used AI for work in 2024 (TALIS 2024) ✓; 72% see AI as a threat to academic integrity ✓. In the US, 83% of K-12 teachers already use generative AI, yet 71% have received no formal AI training source.

The OECD reports strong trial evidence that inexperienced tutors teach better and get better student outcomes when supported by educational GenAI ✓.

• Anthropic, 575,000 student conversations on Claude.ai: ~47% are requests for ready-made answers with minimal cognitive engagement, including direct attempts to evade plagiarism detectors ✓.
• Turnitin: 95% of those surveyed (educators, administrators, students) believe AI is being misused in some way ✓.
• Stanford AI Index 2026: only 6% of teachers find school AI policies to be clear guidance ✓. UNESCO: the technology has outpaced policy debate and regulation ✓.
• An analysis of 124 AI policies at 110 leading universities in the US, Japan and China: not one bans generative AI outright; US universities delegate the decision to instructors, China centralizes, Japan sets institution-level rules source.

The practical conclusion: detectors have lost the race. Credible assessment is shifting to oral and proctored formats and to grading the process — with the AI-interaction log as a new artifact of learning — rather than the final product.

Market-size estimates differ in method but agree on the story — roughly a fivefold rise in five years: Mordor Intelligence puts AI-in-education at $6.9B (2025) → $41B (2030), a 42.8% CAGR; Grand View (May 2026 edition) sees $57.2B by 2033 at 25.9% source. Higher education is the largest segment (44–45%), corporate L&D the fastest-growing (~44.8% CAGR); the fastest-growing application is adaptive assessment (46.7% CAGR) source.

• M&A: ~410 transactions in 2025 (+20% on 2024), 8 IPOs; the markers — Workday → Sana at $1.1B, Coursera → Udemy source.
• Europe took nearly half of global EdTech VC in 2025. Q1 2026: $512M — a slow start source.
• The backdrop: public education spending as a share of global GDP has been falling since 2021; private EdTech money is draining out of classic K-12 into AI tools; capital spending on AI overall is set to double between 2024 and 2026 (Goldman Sachs) ✓.
• HolonIQ's call for 2026: the shift from experimentation to governed deployment source.

UNESCO sets the global frame: a human-centred approach, AI as an accelerator of SDG 4 ✓; AI competency frameworks for students and teachers ✓; the GenAI guidance (2023, 14 languages) with an age threshold and mandatory data protection source; the September 2025 report on learners' rights: 2.6 billion people offline → the risk of an "AI divide" source.

The global marker: PISA 2029 will measure AI literacy for the first time — by 2031 the regulatory debate gets an international metric source.

• By 2030, macrotrends are projected to create ~170 million jobs and displace ~92 million — a net gain of 78 million ✓ WEF.
• Demand for AI literacy skills grew 70% from 2024 to 2025 (LinkedIn) ✓.
• Of every 100 workers worldwide, 59 will need training by 2030; 11 are unlikely to get it — putting their jobs at risk. 39% of existing skill sets will be transformed; 63% of employers name skill gaps as barrier number one source.
• The WEF Four Futures framework makes workforce readiness — the output of education systems — one of the two axes that decide the 2030 labor-market scenario; in the negative "Age of Displacement" scenario, education systems fail to adapt in time ✓.

• Skill erosion. The OECD documents "metacognitive laziness" ✓. Gerlich (2025, n=666): a strong negative correlation between AI use frequency and critical thinking (r≈−0.68), mediated by cognitive offloading; 17–25-year-olds score worst source. Caveat: cross-sectional data — correlation, not causation.
• Inequality → a "metacognitive gap" (the Matthew effect). Unstructured student-facing use will almost certainly widen existing divides source. The counter-evidence: in Utah's RCT (692 schools, 166K students) adaptive reading programs helped disadvantaged groups the most source. The fork: teacher-facing AI narrows gaps; unguided student-facing use widens them.
• Privacy and children's rights. UNESCO: without data protection and accountability, AI puts the right to education itself at risk source. The EU has already banned emotion recognition in classrooms; China bans feeding students' personal data into GenAI.
• The quality of the evidence base. The most-cited meta-analysis of ChatGPT's benefits (Wang & Fan, g=0.867) was retracted on April 22, 2026 — a figure that spread through EdTech slide decks is formally void source. Ask for the primary source behind every effect-size slide.
• Governance failure. Nearly everyone uses AI; 6% find the policies clear ✓. The vacuum is being filled, in practice, by Big Tech product decisions.

The anchor events are already on the calendar: EU AI Act high-risk rules for education — December 2, 2027; the wave of mandatory district AI policies — 2027–28; PISA 2029 → first comparative data around 2030–31.

1. 2026–2028

   From experiments to governed deployment. School systems adopt AI through procurement, audits and oversight; the market for "wild" student-facing AI shrinks under regulation and reputational risk.

2. by 2027–2029

   The general-purpose vs educational AI split becomes law. The OECD's staged approach moves into national policy; certification of pedagogical AI appears — in the EU via high-risk compliance.

3. by 2029

   Teacher copilots become the de facto standard. The economics are hard to argue with: $20 a year against $4,800 for coaching, amid a 44-million teacher shortfall. The main barrier — 71% of teachers with no AI training — feeds a boom in professional development.

4. by 2029

   Assessment is rebuilt before the curriculum is. Adaptive assessment is the fastest-growing segment; the take-home essay effectively dies as a unit of grading in selective systems.

5. by 2030

   AI literacy becomes a measured, mandatory competency. China has introduced it, Alabama made it a graduation requirement, PISA 2029 makes it comparable across countries.

6. by 2030–2031

   A fivefold market and consolidation. $30–45B; point solutions get absorbed into platforms (the Workday–Sana pattern); corporate L&D stays the magnet for capital.

7. 2029–2031

   The first hard longitudinal data shows the differential effect: gains where adoption was structured, eroding fundamentals where it was left to chance. A candidate for the "AI-PISA shock" of the early 2030s.

What will not happen by 2031: mass replacement of teachers (all the evidence points to augmentation), the disappearance of universities, or the closing of the Global South gap (2.6 billion people offline will not get connected in five years).

Ten years in AI is three to four model generations; the honest format is scenarios. The axes follow the WEF framework: pace of AI progress × readiness of education systems.

Structural shifts likely under any scenario

• The education/work boundary dissolves: continuous micro-credentials instead of a degree-for-life; the 59% of workers needing retraining is a permanent condition, not a one-off.
• Higher education comes under business-model pressure: if AI does a junior graduate's job, the degree loses its signaling value — universities drift toward research, socialization, and certifying what AI does not replace.
• The teaching profession is redefined, not eliminated; in poor systems AI becomes the only affordable "tutor" — with all the risks of Scenario B.
• Learning data becomes the most sensitive class of personal data: a longitudinal AI profile of each student — what they know, how they think, where they break down.

The main unknown is the pace of AI itself. If agentic systems reach expert-grade reliability across most cognitive tasks by the early 2030s, the question shifts from "how does AI improve learning" to "what should humans learn in an economy where cognitive work costs less than electricity" — and the goals of education (agency, meaning, community) come to matter more than its methods.

1. AI raises grades and erodes learning at the same time — and that is not a contradiction.

   +48% with AI, −17% without it. Whose metric are we optimizing — the task's or the student's mind?

2. The ban-or-allow debate ended in 2025.

   The real question is who designed your chatbot's pedagogy. +127% against −17% — same model, different wrapper.

3. The best-proven use of AI in education is boring.

   Give teachers back 31% of their prep time. Personalization headlines the panels; the ROI sits in routine work.

4. Academic integrity in its old form is dead.

   47% of student AI queries are requests for ready-made answers. Grade the process, not the product; the AI-interaction log is a new artifact of learning.

5. This is a governance failure, not a technology failure.

   Nearly everyone uses AI; 6% of teachers find the policies clear.

6. AI is the great equalizer and the great divider at once.

   The weakest tutors gain 9 points with AI support, while unguided use sets off a Matthew effect. Implementation design decides which half comes true.

7. The market has already voted — for adults.

   38% of venture deals target workforce training; corporate L&D is the fastest-growing segment. Schools will get AI that was field-tested on corporations.

8. 44 million missing teachers: the strongest argument for AI — and the most dangerous excuse.

   Augmentation, or a cheap substitute for human contact for the poor?

9. The science of AI in education is failing its own stress test.

   The field's most-cited meta-analysis was retracted in April 2026. Ask for the primary source behind every effect-size slide.

10. The world is choosing among three regulatory models: rights, market, state.

    The EU bans emotion recognition in class, China bars open GenAI for primary schoolers, the US is passing 71 state bills. For everyone else, the window to choose closes by 2030.