AI tutoring and personalised learning

Augment-the-teacher beats replace-the-teacher in 2026 deployment

RCT evidence (Carnegie LearnLab, Harvard CRECHE, Stanford CRAFT) shows 0.3–0.6σ effect sizes for autonomous tutoring – meaningful but well below Bloom 1984 2σ human-tutor benchmark. The runaway winner is the teacher-workflow tool: MagicSchool passed 6M+ educators by late 2025, with Brisk Teaching and Khanmigo teacher tools close behind, because they sell back the scarcest classroom resource – teacher time. Khanmigo and Duolingo Max work because they augment existing curriculum, not replace it.

AI tutoring delivers measurable but sub-human-tutor learning gains

Controlled trials show AI tutoring helps, but the rigorous benchmark is 0.3–0.6σ for autonomous tutors – meaningful, not the 2σ of expert 1:1 human tutoring. A carefully scaffolded Harvard physics study (Kestin et al., 2024) reached far higher gains, but only with faculty-built pedagogy and guardrails, on <200 students – it is an upper bound under ideal conditions, not the deployed average.

State of the art (2026)

As of mid-2026 the market has split cleanly into two proven businesses and one unproven one. The teacher-workflow layer is the runaway winner: MagicSchool passed 6 million educators by late 2025, with Brisk Teaching and Khanmigos teacher tools close behind, because they sell back the scarcest resource in a classroom – teacher time. The consumer LLM-tutor layer is validated commercially by Duolingo Max, whose GPT-4-powered Video Call and Roleplay features sustain a high-margin subscription at roughly 168 USD a year even after Explain My Answer went free in January 2026. The unproven layer remains the autonomous classroom tutor: published RCTs still land at 0.3–0.6 sigma, meaningfully short of Blooms 2-sigma human-tutor benchmark. Enterprise demand consolidates around ChatGPT Edu, Claude for Education, Gemini and Microsoft Copilot.

Duolingo Max is validated, but Super is the volume engine

Duolingo Max (GPT-powered Video Call + Roleplay) sustains a high-margin ~168 USD/year tier and is the cleanest commercial proof of LLM tutoring at scale. But it is the premium minority: through H1 2026 Duolingo runs DAU above 40M and MAU above 160M, with Super subscribers outnumbering Max roughly 10:1. Explain My Answer returned to free in January 2026, narrowing Super–Max differentiation. The economics-positive signal is real; the headline conversion is Super, not Max.

Teacher augmentation, not replacement

The winning deployment model is AI as a tireless teaching assistant that handles practice, formative assessment, and personalised feedback, freeing teachers for higher-order instruction, mentorship, and social-emotional support. Districts that frame AI as teacher replacement face union resistance and poorer outcomes.

Enterprise LMS consolidates around four hyperscaler-aligned platforms

OpenAI ChatGPT Edu, Anthropic Claude for Education, Microsoft Copilot for Education (with Khanmigo integration), Google Workspace + Gemini define the K-12 + higher-ed enterprise market. Independent pure-plays (Synthesis, Eedi, Carnegie Learning) survive in vertical niches but the platform economics consolidate around the four. ChatGPT Edu universities signed list (ASU, Wharton, Caltech, Texas A&M, Columbia, Oxford) is the reference.