AI Tutoring in 2026: What the Research Really Shows

A Rapid Shift From Novelty to Infrastructure

By 2026, AI tutoring has stopped being a curiosity bolted onto homework apps and has become a standard feature of how schools and students approach studying. Google has folded its education-tuned LearnLM models into Gemini, offering a "Guided Learning" mode that walks students through study plans, quizzes, flashcards and visual explanations rather than simply handing over answers 6. In June 2026 the company added adaptive "study notebooks" that run a diagnostic quiz, then continuously update a personalized lesson plan as a student's performance changes 7. Google has also rolled these capabilities into Gemini for Education and Google Classroom, pairing them with administrative controls, data protections and teacher-led activities so schools can manage how the tools are used 818. Elsewhere, iAsk AI has repositioned its search engine as an education hub offering tutoring and guided homework help 5, while ed-tech commentators describe a landscape in which AI has moved from a supplementary classroom tool to a central engine of instruction 1.

Khan Academy's Khanmigo represents a narrower, more deliberately constrained approach. Rather than acting as a general-purpose answer generator, it is designed to coach students toward a solution, particularly in math, without revealing it outright 39. That design choice — coaching versus answering — has become the fault line running through nearly all of the credible research on whether AI tutoring actually improves learning.

What the Strongest Trials Actually Found

The most substantial evidence to date comes from a two-year cluster-randomized trial across 18 middle schools in Hamilton County, Tennessee, involving 6,902 student-term observations 910. Students assigned to use Khan Academy with Khanmigo during their daily remedial math block gained about 1.26 national percentile ranks per term, translating to roughly 0.06 to 0.08 standard deviations per school year, with the effect rising to an implied 0.14 standard deviations for students who participated actively for a full year 910. Crucially, the researchers found these gains resembled what Khan Academy's practice platform produces without any AI tutor at all, suggesting the effect was driven mainly by an extra 30 minutes of weekly practice rather than by anything uniquely "intelligent" about the chatbot 910.

The usage data explains why. Ninety-six percent of students tried Khanmigo at least once, but engagement was shallow: the median student messaged it in only about a third of practice days and 14 percent of exercise sessions, and just 14.5 percent of all messages contained an actual mathematical question or reasoning step, while nearly 40 percent were simply bare answers typed into the chat 10. Companion experiments in the same research program — testing mastery-based progression and a virtual home-tutoring offer — found similarly modest and often statistically inconclusive effects once take-up and dosage were accounted for, reinforcing that engagement, not model capability, is the binding constraint 10.

A very different picture emerges from a Harvard-led randomized trial comparing a custom AI physics tutor to an active-learning classroom. Students using the AI tutor posted median post-test scores of 4.5 versus 3.5 for classroom peers, more than doubling learning gains relative to baseline, while also reporting higher engagement and motivation and finishing in less time 111217. That result is striking, but it involved a purpose-built tutor engineered around specific pedagogical principles rather than open-ended chatbot access, and a review of the study cautioned that its dramatic effect size may reflect the weakness of the classroom comparison rather than absolute AI superiority 12.

Human Oversight Keeps Outperforming AI Alone

Several of the more rigorous 2026 studies point toward hybrid models — AI supporting a human tutor — as the most consistently effective approach. Stanford's Tutor CoPilot trial, run with more than 700 tutors and over 1,000 students from underserved communities, found that tutors given real-time AI guidance saw their students become four percentage points more likely to master a session's topic, with gains reaching nine percentage points for students paired with lower-rated tutors, at a reported cost of about $20 per tutor per year 1319. Analysis of more than 350,000 messages showed the tool nudged tutors toward asking students to explain their reasoning rather than offering generic encouragement, though the study did not detect a significant bump in end-of-year test scores over its two-month window 13.

Google's LearnLM experiment used a similar supervised structure. In a trial with 165 UK secondary students, expert tutors approved 76.4 percent of the model's proposed messages with little or no editing, and students who received LearnLM-supported tutoring solved subsequent novel problems at a 66.2 percent rate, compared with roughly 61 percent for students tutored by humans alone and 56.2 percent for those given only static hints 1419. Researchers described this as evidence that supervised AI can match or exceed unaided human tutoring on transfer to new material, even if it isn't a wholesale substitute for one 14.

A Stanford SCALE briefing pulled these threads together bluntly: nearly half of students given access to an AI tutor simply didn't use it, and the emotional persistence a human tutor provides remains hard to replace 15. Its message to school leaders facing budget pressure was that AI works best layered onto human tutoring, not swapped in for it 15. Smaller studies reinforce the caution: an undergraduate physics experiment using Khanmigo against a Google search comparison group found no statistically significant difference in learning outcomes between conditions, even though students said they liked Khanmigo's step-by-step guidance 16.

Why Coverage of the Numbers Diverges So Sharply

Different outlets are, in effect, measuring different things. Vendor and tech-press framing tends to emphasize adoption and product breadth — Gemini as a multimodal "learning companion," Khanmigo as a Socratic math coach, iAsk AI as an emerging hub 135678. Peer-reviewed and working-paper research is far more conservative, stressing sample sizes, study duration and the gap between short-term accuracy and durable, transferable learning 9101314. Meanwhile, some aggregator sites cite eye-catching statistics — claims of 54 percent higher test scores, 70 percent higher course completion, or exam scores rising by wide margins — that should be read with skepticism, since they often compile secondary or vendor-supplied figures rather than describing a single controlled study 20. Broader meta-analyses of intelligent tutoring systems sit in between: they generally find positive effects relative to traditional instruction, but with effect sizes that shrink once AI systems are compared against other technology-assisted or non-intelligent tutoring rather than an unaided classroom 17.

The Academic Integrity Problem Grows Alongside the Tutoring Boom

The same technology capable of coaching students through a problem can just as easily produce a finished essay or solved homework set, and that duality is straining how schools define legitimate work. Data cited in industry statistics show AI text generation among UK undergraduates more than doubling in a single year, from 30 percent to 64 percent, turning academic integrity from a background worry into a pressing institutional issue 20. Even so, the same data suggest most student AI use is defensible: over half of students report using AI mainly to get explanations or find information, and a majority say they treat it as a tutor rather than a shortcut for cheating, even as a meaningful minority — roughly a fifth — admit to submitting AI work without disclosure despite knowing it's against the rules 20.

Schools are responding less by trying to detect AI use after the fact and more by rethinking what counts as evidence of learning. Coverage of the classroom rollout notes that districts are simultaneously chasing the promise of personalized instruction and wrestling with concerns about diminished human connection and the fairness of enforcement 34. U.S. Education Secretary Linda McMahon has voiced support for AI's expanding classroom role even as commentators warn that younger students, in particular, are becoming de facto test subjects for still-unproven tools 4.

What It Means Going Forward

Taken together, the 2026 evidence supports a more measured story than either AI boosters or skeptics tend to tell. Purpose-built, pedagogically constrained tutors can produce meaningful gains, especially when paired with human oversight, as seen with Tutor CoPilot and LearnLM 131419. Unstructured or lightly used AI access, as in the Khanmigo middle-school trial, tends to produce modest gains that track added practice time rather than any dramatic leap in comprehension 910. And the sharpest classroom debates are no longer only about whether AI tutors work, but about how to verify that what a student turns in reflects what they actually learned — a question the education system is still working out school by school, assignment by assignment 3420.

The throughline across nearly every study and news account is the same: access to an AI tutor is not the same as learning from one. Engagement, supervision and the design choices that separate a coach from an answer machine will determine whether the AI tutoring boom of 2026 becomes a durable improvement in how students learn or simply a faster way to arrive at the same old problem of unequal effort and attention.