Education and personalized learning

📊Evidence & Reach

Khan Academy's Khanmigo (built on GPT-4) gives every Khan Academy user a free patient AI learning partner that asks questions instead of giving answers, and peer-reviewed early results suggest measurable gains for students who use it consistently. Carnegie Learning's MATHia adaptive platform shows similar gains across thousands of US school districts. Duolingo's AI conversation features (Duolingo Max) enable real-time language practice with native-pace responses. Meta's No Language Left Behind project supports 200+ languages, many getting machine-translation support for the first time. The efficacy evidence now reaches past school-age classrooms into professional training: a 2026 Stanford Law study ran nearly three thousand blind comparisons and found AI tutors out-performed law professors in about three-quarters of contract-law matchups. The largest test of the general case is more demanding, and it names the condition this whole area depends on. The OECD's PISA 2025, published September 8, 2026 and covering more than 760,000 fifteen-year-olds across 91 countries, found that students who never use AI generally outperformed those who do in science once socioeconomic status is controlled for. But the effect tracked *use*, not access: summarising an assigned reading or drafting an essay tracked worst, general help learning tracked far better, and students who had been taught to judge the quality of what an AI tells them outperformed non-users outright. That last group is the one this page is about. The promise here is real and it is conditional — a patient tutor helps when it makes a student do the thinking, and hurts when it does the thinking for them.

Sources:Khan Academy Khanmigo researchBloom's 2-sigma problem + AI learning (Stanford)Carnegie Learning impact researchAI outperforms law professors (Stanford Law, 2026)PISA 2025 Results (OECD, September 2026)