AI tutors in 2026: what trials show and what comes next

Vladimir Arustamov 6 Nov 2026 7 min read

Guides

AI tutors can teach, but most students don't use them enough to benefit, and that is the main finding of the 2026 research. Designs that start the interaction themselves, such as scheduled sessions and guided lessons, look stronger than a chat box that waits for a question.

In the Digital Education Council's 2026 survey of 45,398 students and faculty, 88% of students said they use AI in their learning. Yet when the University of Maryland added a GPT-4o study assistant to course sites in fall 2025, only about 15% of the students offered it used it even once. They also had ChatGPT and Gemini through the university.

This guide reviews nine randomized trials and the tech behind the main products, with figures taken from the studies, company posts and public filings and checked on November 5, 2026. The trials differ in age, subject and length, so their results can't be compared directly.

AI tutors and student engagement

Access alone rarely solves the problem, even when the setup looks close to ideal. In a two-year trial in 18 Tennessee middle schools, the AI tutor was free, built into mandatory math sessions and set up to coach instead of giving answers. Scores rose by about 0.06 to 0.08 standard deviations a year, roughly what the same platform delivered without AI.

One likely reason is how little students used it: the median student messaged the tutor on a third of practice days and in 17% of the sessions with a mistake. Making the tutor open by default in January 2026 produced no detectable change.

In a Stanford trial with 355 elementary students, adding a human tutor raised AI platform use by only one to about four minutes a week, and reading didn't improve.

Our read: asking for help costs more than product teams assume. A stuck student has to notice the confusion, admit it and put it into words.

Students who do use chat can hit the opposite problem. In a Turkish high school, plain GPT-4 lifted practice grades by 48%, then left students 17% below the control group on an unassisted exam. At Maryland, 73.8% of students who used the assistant asked for information, explanations or solutions.

AI tutoring that works

If students rarely ask for help, the obvious fix is to stop waiting for them to ask, and the better results in the trials tend to come from systems that take the first move. A custom GPT-4 tutor at Harvard produced more than double the median learning gain of an active-learning class in a 2023 trial with 194 students.

Eedi, a UK math platform, opens the chat itself once a diagnostic question exposes a misconception. In a 2025 trial with 165 students, Google's LearnLM drafted the messages and expert tutors reviewed them before they went out. Students then solved novel problems 66.2% of the time, against 60.7% with human tutors alone. Google and Eedi wrote the study.

A larger follow-up with 1,179 students in ten UK schools was less clear-cut: supervised AI tutoring matched human tutoring on in-platform measures, but showed no clear advantage over standard support on a broader standardized math test.

Two cheaper designs hand the first move to someone else. In Stanford's Tutor CoPilot it belongs to a human tutor, who gets AI guidance during live sessions; that made students 4 percentage points more likely to master a topic, at about $20 per tutor per year. Rori leaves it to the timetable, with two scheduled math sessions a week over WhatsApp. Its Ghana trial reported an effect size of about 0.37, which the authors say covers only the first year.

Native tools such as NativeLingo go a step further and remove the lesson altogether. They rely on native learning, which builds practice into tools people already use, such as web pages, video subtitles or homework search. NativeLingo starts practice on the page or in the subtitles a learner already has open.

AI language tutors

Paying adults hold back too. Speak's testing found that interrupting a tutor doesn't come naturally to many learners, so the company built in moments where the tutor checks understanding and asks for an answer in the learner's own words. Speak moved to direct speech-to-speech on OpenAI's Realtime API in 2024, and on September 10, 2026, it began a limited rollout of lessons on GPT-Live-1, a model that listens while it speaks.

Duolingo has published the clearest account of how such a tutor is put together. For Video Call, learning designers write the system instructions, and the first question is generated separately because a single large prompt produced overly complex sentences. After each call, the transcript goes back to the model to extract facts for the next conversation.

The price of talking is falling fast. Duolingo's CEO said in August that a video call now costs under $0.01, down from about $0.30, largely thanks to open models, and OpenAI lists GPT-Live-1's voice layer at $0.05 per minute. When any app can rent the same voice, the voice stops setting products apart, and what keeps people coming back every day matters more. Duolingo reported 64.9 million daily active users in Q3 2026, up 29%.

For a wider comparison of voice products, see our guide to the best AI voice assistant.

Spaced repetition and word swaps

Cheap conversation doesn't help much if learners forget what they practiced, and some useful learning software keeps language models out of the critical path to focus on memory instead. FSRS, the open-source scheduler Anki added in version 23.10, tracks difficulty, stability and retrievability for each card and fits them to a learner's review history.

The method rests on research like Karpicke and Roediger's 2008 experiment, in which students who kept being tested recalled about 80% of word pairs a week later, against 36% for those who kept studying without further tests and 33% for those who dropped the words.

Recognizing a word and producing it are different skills. In a 2010 study of 20 advanced readers, words met more than ten times in a novel were recognized 84% of the time but recalled only 55% of the time.

AI tutor tech stack

Across the products covered above, the clearest difference is who starts the turn.

Project	Who starts the turn	Publicly known stack
Khanmigo	The student, opens by default since 2026	LLM text chat, launched with early GPT-4 access
University of Maryland VSA	The student	GPT-4o with RAG over course materials
LearnLM on Eedi	The platform, after a diagnostic error	Pedagogy-tuned model with human review
Tutor CoPilot	The human tutor	LLM guidance during live sessions
Rori	The timetable	WhatsApp chat tutor
Speak	The lesson script	Realtime API, now GPT-Live-1
Duolingo	The daily habit	Birdbrain, structured prompts, growing use of open models
Anki with FSRS	The scheduler	Open-source memory model
NativeLingo	The page or subtitle, then a review	On-device dictionary, UniMorph and UD tables, FSRS

The tools differ in who starts the interaction: the student, the tutor, or the platform.

The future of AI tutors: forecasts to 2027 and 2030

HolonIQ counted $1 billion of edtech venture funding in the first half of 2026, down 26% year over year, in a half-year when global venture funding hit a record $510 billion. Chegg's revenue fell 47% in the first nine months of 2026. Basic explanation is now cheap: ChatGPT, Gemini and Claude all rolled out learning modes within weeks of each other in 2025.

Our near-term forecasts are predictions, not confirmed plans:

By the end of 2027, student tutors will move out of the side panel and into the exercise itself, because default placement alone didn't change behavior in Tennessee.
By the 2027–28 school year, at least one large US district will pay for measured use instead of seats. One Florida district's AI tutoring program cost about $2.5 million in its first year, half of it paid by a foundation.
By the end of 2027, live conversation alone won't justify a premium subscription, because any startup can rent the same voice model.

Further out, we expect a bigger shift by 2030: most everyday AI learning will happen inside tools people already use, with no tutor to open and no lesson to book. With explanations nearly free and a conversation costing less than a cent, what stays scarce is the learner's effort to start.

Google has already rolled out Guided Learning in Classroom globally and plans to pilot guided homework help inside Search for US schools later in November. McGraw Hill is connecting its course content to Microsoft 365 Copilot, with a limited beta planned for early 2027.

Language learning is one of the first places where this shift is visible. A small group of browser tools now brings practice into everyday reading and video, and NativeLingo is a good example of where that market is heading: a few new words on the pages people read and in the YouTube and Netflix subtitles they watch, followed by a typed review spaced with FSRS.

Dmitry Ratner, NativeLingo's CEO, links this design to the barrier the trials keep running into.

Starting is the hardest part of learning a language, so we put practice inside what people already read and watch, then have them type each word from memory.

The product has already found its users among learners who keep dropping lesson apps but still read and watch online every day. Its effect on long-term recall still needs independent measurement.

FAQ

Do AI tutors work?
They can. Trials at Harvard and on Eedi found strong results, but outcomes depend on design and on how much students use the tutor.

Why do students avoid AI tutors?
The 2026 trials point to low initiative. Many students don't ask for help even when a tutor is free and one click away.