AI tutors are not new. I built one in the late 1990s that used an ideal model of competence and tracked student performance, recommending what to do next so that the student model moved towards the ideal model. I also spent several years involved in a tutoring system used across Higher Education, including an entire Biology degree at a major US university. Personalised learning was around long before ChatGPT.
What has changed is the technology. AI tutors are now more capable, conversational, adaptive and dramatically cheaper. The question is no longer whether AI can tutor. It can. The more interesting question is how much human involvement is actually necessary.
The research is starting to provide some surprising answers.
AI teaching assistant that fooled everyone
Back in 2016, Ashok Goel at Georgia Tech introduced Jill Watson, an AI teaching assistant that answered student questions and dealt with routine queries. Students thought Jill was human. One became suspicious because the replies arrived too quickly. The solution? Slow the AI down to make it look more human!
Students even nominated Jill for a teaching award. Why? It was available 24/7, consistent, useful and never tired of answering the same questions. More recent versions use generative AI and course materials to provide increasingly sophisticated support.
This was an early demonstration of AI's ability to take on routine teaching support. But the strongest evidence for AI tutoring now goes well beyond answering questions.
AI tutors with teachers
In Nigeria, De Simone et al. (2025) ran a randomised trial in which students worked in pairs with AI across twelve 90-minute sessions. Carefully designed prompts encouraged explanation, practice and revision rather than simply supplying answers. The results were impressive. Students gained showed improvement across English, AI knowledge and digital skills, the researchers equated to 1.5–2 years of normal schooling progress.
Teachers were involved here. They introduced topics, provided prompts, monitored progress and encouraged reflection. This was structured AI tutoring, not simply letting pupils loose with ChatGPT.
Something similar happened in Ghana. Henkel et al. (2024) evaluated Rori, an AI mathematics tutor used across 11 schools. Pupils completed two 30-minute sessions each week over eight months, with gains described as equivalent to an additional year of learning. They worked individually with the AI, but within timetabled, again in supervised teacher-present school sessions.
In an interesting variation, and move towards seeing how well AI tutors perform, Google and Eedi's UK research with LearnLM went a step further. AI generated mathematics tutoring messages, which human tutors checked before sending to pupils. The AI-supported tutoring performed at least as well as human tutoring on measured outcomes, and slightly better on novel problem-solving. Remarkably, 76.4% of the AI's messages needed no changes or only minor edits.
These studies suggest that AI can do much of the intellectual work of tutoring, humans providing structure, oversight and intervention.
But do those humans need to be qualified teachers?
Guides, not teachers
In Uttar Pradesh, India, an earlier attempt to use Khan Academy without human support had failed to deliver the hoped-for results. Researchers then tried a different approach across 83 government middle schools.
They introduced lab-in-charges, not qualified mathematics teachers but facilitators who organised sessions, assigned activities, helped with the software and kept pupils on task. Weekly practice rose from 7.2 to 47.4 minutes. Mathematics scores improved by almost half a standard deviation.
The distinction matters. These adults were not teaching mathematics in the conventional sense. They were making sure pupils actually used the technology productively. Of course, their contribution was part of a larger intervention, so we cannot isolate its precise effect.
Alpha School
Alpha School has put the ‘guide not teacher’ model to the test in the real world, on scale. Alpha claims pupils can complete their core academic learning in just two hours a day using adaptive software and AI, leaving afternoons free for sport, creative activities, coding and other interests.
Traditional teachers are largely replaced by Guides, whose role is to motivate, monitor and coach. I have spoken and debated with one of these Guides. The distinction was clear. The job was not to teach subjects but to get pupils to learn and keep them on task.
Alpha is highly metrics-driven, with mastery thresholds, progress tracking and even controversial token rewards. Its internal figures from Brownsville, Texas, claim dramatic gains, including second-grade mathematics scores rising from the 31st to the 84th percentile in a year. Impressive claims, but not independently validated research. With high fees, a selective intake and motivated parents, it is difficult to know how much of the success comes from the technology. But this is a promising and innovative experiment, with paying parents.
The idea of independent learning with adult guidance is not revolutionary. Montessori has been doing something similar for over a century, with learner choice, structured materials and close adult observation. Unlike many recent experiments, it has a substantial research base. Randolph et al. (2023) found positive academic outcomes, while Lillard (2025) reported better kindergarten outcomes in a large US randomised study.
Rocketship Public Schools combined teaching with adaptive software and achieved strong results. But its original large Learning Labs created behavioural and organisational problems, forcing a redesign around smaller groups and closer teacher involvement.
Then came AltSchool. Founded in 2013 by former Google executive Max Ventilla, it raised around $175 million to create highly personalised, technology-rich schools. It attracted enormous publicity but struggled with costs, staffing and scaling. By 2019 it had stopped operating schools.
This is the challenge for Alpha. A model may work with selected pupils and highly committed parents, yet fail to scale across ordinary schools.
The real test is not whether a few expensive schools succeed, but whether the model works for everyone else.
What happens when we remove the humans?
So far, the evidence suggests teachers may be necessary, for now, with very young learners. But what about older school students, college students and adults, who are more capable of managing their own learning?
A Harvard study by Kestin et al. (2025) compared AI tutoring with active classroom learning in undergraduate physics. Using a randomised crossover design, students experienced both approaches. The AI tutor produced more than twice the learning gains, in less time, with students reporting greater engagement and motivation. This wasn't a general chatbot. It was a carefully designed tutor using structured explanations, feedback and scaffolding. Nevertheless, students learned more without a teacher delivering the lesson.
That's significant because the comparison was with active learning, not a passive lecture. The implications become even more interesting as general-purpose AI models improve.
Lee et al.'s (2026) Thinking-KT shows how reasoning models can use a learner's previous answers, errors and patterns to predict performance, identify weaknesses and recommend what to study next, without specialised fine-tuning. This takes AI beyond answering questions. It starts to build a developing model of the learner, adapting feedback and practice accordingly.
It's the direction I anticipated in my book AI and Learning (2024), where I described, in detail, the possibility of a 'Universal Tutor'.
Language learning is already demonstrating the value of relatively autonomous AI tutoring. You can practise conversation, pronunciation, listening and grammar whenever you want, at your own level, without embarrassment. The AI provides immediate feedback and endless opportunities to repeat and improve.
Xu and Wang's (2024) meta-analysis of 40 studies found significant improvements in English language learning achievement. Du and Daniel's (2024) review of 24 studies reported benefits in speaking practice, pronunciation, confidence, motivation and reduced anxiety. Much of this involves learners using AI informally, without constant supervision. The broader evidence is also encouraging. Gençdal and Çoğaltay's (2026) meta-analysis of 215 effect sizes found a large positive average effect on academic achievement (0.80) and a moderate effect on motivation (0.64). These are averages across varied interventions, not proof that every chatbot is a good tutor. Pedagogical design remains important.
It is increasingly difficult to argue that effective learning always requires a human teacher. It has always been clear that teaching is not always a necessary condition for learning. Most learning in the workplace and life takes place in the absence of human teachers.
Cost changes everything
Perhaps the most disruptive evidence comes from Northcutt et al. (2026), who compared AI and expert human tutoring in verbal reasoning, quantitative reasoning and analytical writing.
The AI tutors produced statistically equivalent short-term learning gains. Then came the costs. An hour with an expert human tutor cost $75. The reported AI tutoring cost was just seven cents. That's around thousandfold difference in hourly cost. The comparison doesn't include every possible development, deployment or supervision expense, and short-term results don't settle questions about long-term learning. But the economics are extraordinary.
For centuries, one-to-one tutoring has been among the most effective, yet expensive, forms of education. It has largely been available to those who could afford it. AI makes it potentially available to everyone. If AI tutoring can deliver substantial learning gains at a fraction of the cost, the potential to reduce educational inequality is enormous.
Do we still need teachers?
The evidence points towards a distinction between tutoring and teaching.
AI is increasingly capable of explaining, questioning, diagnosing, providing feedback and adapting practice. These are central components of tutoring, and it can already perform many of them effectively. What humans often contribute is something different: motivation, supervision, social interaction, safeguarding, judgement and organisation.
For younger children, that human presence appears valuable. But the Indian evidence suggests it need not always come from a qualified subject teacher. For older learners, university students and adults, the need for supervision may be substantially lower.
This doesn't mean teachers disappear. It means we should stop assuming that every educational task requires a teacher.
Nor should we assume that putting ChatGPT in front of pupils automatically improves learning. The most convincing results come from carefully designed tutoring experiences, with appropriate scaffolding, feedback and opportunities for practice.
Alpha School may be pointing towards a different model, although its claims need independent validation. Harvard's physics study and developments such as Thinking-KT suggest that increasingly capable AI tutors may also work effectively without human supervision.
The likely future is not one model but several: teacher-supported AI tutoring, guide-supported tutoring and autonomous AI tutoring, depending on the learner, subject and context.
There is, however, a more fundamental point. We have spent centuries organising education around the scarcity of teachers and tutors. AI is removing that scarcity. Classrooms, timetables, age-based cohorts, lectures and fixed courses are, in part, ways of distributing limited human expertise among large numbers of learners. When personalised, one-to-one tutoring becomes abundant and extraordinarily cheap, those assumptions become open to question.
The real challenge is not simply how to fit AI into our existing educational institutions. It is whether those institutions should continue to look the way they do.
