Stanford review finds strongest AI tutoring results come from tools that support human tutors

Research from the National Student Support Accelerator and AI Hub for Education finds AI -only tutoring still has major evidence gaps, with some students barely using the tools at all New Stanford research finds AI tutoring currently has the strongest evidence when it supports human tutors rather than replacing them AI tutoring is moving into schools faster than research can establish what works, and the strongest evidence so far favors using artificial intelligence to support human tutors rather than replace them, according to a new Stanford University research review. The National Student Support Accelerator and AI Hub for Education at Stanford’s SCALE Initiative examined research on high-impact tutoring and emerging AI-enabled models, alongside interviews with tutoring providers, platform developers and researchers. Published in August 2026, AI Tutoring is Not a Monolith: What We Actually Know separates AI tutoring into different models according to how much human involvement they retain. Its central finding is clear: live, human-led tutoring remains the model backed by the strongest evidence. AI can help tutors work more effectively, but fully automated tutoring has not yet built a comparable evidence base. The term “AI tutor” can cover very different experiences, from an AI system quietly helping a human tutor during a live session to a student working alone with software. The researchers argue that those models should not be treated as interchangeable. High-impact tutoring itself already has a substantial research base. The brief notes that students receiving it have shown learning gains ranging from three to more than 15 months across different grades and subjects, with consistent tutors, regular sessions, small groups and strong student-tutor relationships among the features associated with effective programs. The question is how much of that survives when AI takes over more of the teaching. AI-only tutoring has an engagement problem The weakest evidence currently sits at the most automated end of the spectrum. Two randomized controlled trials across two school districts found that when students were left to use an AI tutoring platform independently, between 40% and 47% never used it at all. Even among those who did engage, usage remained low. Students used the platform for only four to five weeks across a multi-month intervention, averaging around two to five minutes per week. That was well below the level of participation associated with reading gains. Adding human check-ins and encouragement helped students engage with the AI platform, but did not produce an improvement in reading achievement in those trials. The researchers stopped short of concluding that AI-led tutoring improves student outcomes. They describe its effectiveness as unknown and point to unresolved problems around engagement and the amount of time students actually spend using the technology. A much larger study involving 181,000 students using supplemental math software points to the same implementation challenge. Only 5% reached the recommended 30 minutes of weekly use, while 41% never logged in. Teacher, school and district factors accounted for 57% of the variation in usage. In practice, having an AI tutor available is not the same as students consistently using it. The brief argues that dosage remains critical. High-impact tutoring embedded within the school day typically involves three or more sessions each week, totaling around 90 minutes for at least 10 weeks. One tutoring provider interviewed for the research puts it simply: “If teachers are not bought into using [AI] in their classroom practice or setting homework, then there is no impact of [AI], because dosage is absolutely key.” AI performs differently when the tutor stays in charge The picture changes when artificial intelligence supports a human tutor rather than taking over the session. Research highlighted in the review found that students whose instructors used an AI coaching system providing real-time recommendations were four percentage points more likely to master lesson topics. For students taught by lower-rated tutors, the increase reached nine percentage points. That model keeps the tutor responsible for the relationship and the teaching while AI works in the background. The Stanford team identifies a range of ways AI could play that supporting role, including giving tutors real-time pedagogical suggestions, generating curriculum-aligned practice materials, analyzing student progress, identifying misconceptions, helping group students according to current needs and reducing administrative work. A system offering one-to-one AI interaction does not automatically inherit the evidence behind one-to-one human tutoring simply because the ratio is the same. The review argues that the benefits associated with small tutoring ratios also come from consistent engagement, personalized instruction, the ability to spot misunderstandings and responsiveness to student cues. It also cautions against assuming individual AI tuition is always better than a small human-led group. In secondary mathematics, for example, research cited in the brief indicates that peer collaboration can itself contribute to positive outcomes. The researchers describe human relationships as a defining part of high-impact tutoring rather than an optional extra. One human-led tutoring provider developing AI delivery tells the researchers: “You're not necessarily replicating that experience [of tutoring with humans]. I think it's a mistake to try and—you know, an AI is not human. You should focus on the strengths of it [AI].” Schools urged to separate student-facing AI from tutor tools The review ultimately gives states, districts and schools a more specific test for AI tutoring than asking whether “AI tutoring works.” It recommends separating direct-to-student systems from tools designed to support teachers and tutors, because the evidence behind the two approaches is not the same. For now, the researchers recommend prioritizing uses that expand human-led teaching capacity, including tutor coaching, lesson preparation, curriculum-aligned materials, attendance and dosage monitoring, and analysis of student data. The brief also calls for human review of AI-generated instructional materials before they reach students. Privacy is treated as a baseline requirement. The researchers say schools and districts should have enterprise-grade data agreements and training around personally identifiable information in place before deploying AI tutoring tools. There are also wider questions still to answer. The review notes unresolved concerns around student safety, long-term cognitive development and the growing use of AI systems for emotional connection. Its overall position is therefore neither that AI has no place in tutoring nor that automated tutors are ready to replace people. The evidence available today supports a narrower role: using AI to increase tutor effectiveness and educator capacity while keeping human relationships at the center of high-impact tutoring. As one tutoring provider interviewed for the research puts it: “I don't think it's a thing to be rushed. AI is exploding everywhere, but I think this is one where being thoughtful and taking our time to do it the right way is going to be the right answer.” The brief was written and edited by Chayne Turano, Valeria Pihl, Chris Agnew, Lauren Ziegler and Susanna Loeb of Stanford University’s SCALE Initiative.