Alpha School Is Growing Fast. The Evidence for AI Tutors Isn't.
A Penn State review of the research finds no evidence AI tutors outperform human teachers, even as tuition-funded AI schools expand rapidly.
The pitch is seductive. The evidence isn't there yet.
Alpha School, the Austin-founded chain that promises two-hour academic days powered by AI tutoring software, has grown from a single campus to more than 15 locations across the United States, including New York and San Francisco, with annual tuition running $40,000 to $75,000. Unbound Academy and Khan Lab School have followed similar models, betting that adaptive software can do more with less classroom time than a teacher standing at the front of a room.
None of it has yet been shown to work better than what teachers already do. That is the conclusion of a research review published this month by Gerald K. LeTendre, a professor of educational administration at Penn State, writing for The Conversation. His argument is not that AI tutoring is useless. It is that the evidence base does not support the claim increasingly used to sell it: that AI tutors are superior to, or a viable replacement for, human instruction.
What the studies actually say
LeTendre's review pulls together several strands of research, and none of them lands where the AI-school marketing suggests.
A 2020 review by the National Bureau of Economic Research found that human tutoring, across a wide range of formats and subjects, produces consistent learning gains. That remains the baseline against which newer technology gets measured, and it is a high bar.
Comparisons between computer-based and human tutoring from 2025 found no significant difference in learning outcomes between the two formats. A separate body of 2025 research on AI tutoring specifically found positive effects across subjects and grade levels, and a Harvard study of an AI physics tutor reported students learning faster and with more motivation. A 2026 Brookings Institution analysis noted that generative AI has genuinely improved computer-assisted tutoring by letting students interact in ordinary language and by adapting sessions to a student's progress in real time.
Taken together, that is a case for AI tutoring as a useful supplement, not a substitute. The strongest results in LeTendre's review come from combination models: a 2024 study of middle schools in low-income areas found gains when human tutors had AI tools to support their sessions, not when AI replaced them. A 2026 study on lesson planning found something similar in a different context, that experienced teachers who use AI to draft materials tend to critically revise the output and tie it back to their curriculum, rather than deploying it unchanged.
The pattern across the research is consistent: AI helps most when a teacher is still in the loop, deciding what to keep, what to cut, and what a specific student needs.
LeTendre calls the replace-or-supplement framing a false dichotomy. Tech leaders promoting AI schools have an obvious incentive to sell the replacement story. The data, so far, backs the supplement story.
Why this matters beyond one Penn State review
The findings echo what England's own evidence-gathering has turned up. Ofsted's study of 21 early-adopter schools and further education colleges, commissioned by the Department for Education, found that leaders who had been using AI for at least a year overwhelmingly deployed it to support staff, not to replace them, with time saved on lesson planning, resource creation and admin redirected toward teaching itself. One college principal told Ofsted's researchers that AI adoption feels like the Wild West, and that leadership's role is simply to police it responsibly. That is a description of cautious management, not a wholesale handover of instruction to software.
The DfE's own guidance, most recently updated in May 2026, makes the same distinction explicit: there is reasonable evidence that AI reduces teacher workload on administrative tasks, and comparatively little evidence yet on the impact of pupil-facing AI tools used directly for instruction. That gap is precisely what LeTendre's review is describing at a larger scale, across a wider set of studies, most of them American.
For UK schools, none of this is an argument against AI in the classroom. It is an argument against buying the version of the pitch that treats AI as a teacher substitute, whether that comes from a US microschool chain or from a vendor demo closer to home. The research consistently rewards the model where a teacher decides how AI output gets used, not the model where AI runs the lesson.
What to do
Treat any tool or programme claiming to replace teacher-led instruction, rather than support it, as an unproven claim, not a settled one. Where AI is already used for tutoring or personalised practice, keep a teacher reviewing and adapting the output, in line with what the 2024 and 2026 studies found actually produces gains. If your school is piloting adaptive learning software, ask what evidence the vendor can show beyond usage figures and satisfaction surveys, and treat NBER-style outcome data as the standard, not marketing copy.
What to watch
Alpha School's expansion into new cities will keep generating headlines, and its backers will keep making replacement claims that outrun the data. Watch instead for independent, peer-reviewed outcome studies on AI-heavy school models as more of them reach the age where results are measurable, and for whether Ofsted revisits its early-adopter cohort now that many of those schools have had AI tools in place for over two years.