AI Tutoring Evidence Splits as Schools Tighten ChatGPT Rules

A semester of mixed signals

This fall, three questions about AI in education are pulling against each other. Does AI tutoring work? How much access should students get to tools like ChatGPT? And can schools still tell what students learned on their own? The research and the policy moves of the past few months don't agree. Read together, though, they lead to one conclusion: the tool itself matters less than how it is built into teaching. And most schools haven't yet worked out how to build it in well.

The newest evidence is the most uncomfortable. A randomized controlled trial released in October 2026 gave some sections of courses at a large U.S. public university access to a generative AI tutor built into the course. The study covered 2,379 undergraduates and 30 instructors across several disciplines. Students with access to the tutor finished with final grades 0.27 to 0.37 standard deviations lower than peers without it.24 Their activity on the course's learning platform fell even more sharply. Participation in activities the instructors had designed dropped by 0.90 standard deviations.24 First-generation students lost more than twice as much in grades as students whose parents had attended college.24 The authors conclude that the tutor may have pulled students away from coursework that was already doing its job, without adding anything that made up for it.24

The case for AI tutors is real, but narrow

The October trial doesn't settle the question, because the earlier randomized studies point the other way. In June 2025, Harvard researchers published a study in Scientific Reports comparing a custom physics tutor with in-class active learning. They reported effect sizes of 0.73 to 1.3 standard deviations, and the AI group spent a median of 49 minutes on the material versus 60 for students in class.27 The study used 194 university students, and each student tried both approaches, which strengthens the comparison but limits its scale.23 Even coverage that celebrated the result describes the tutor as a complement to classroom teaching, not a replacement for it.23

In K-12, two randomized trials reviewed by Stanford's National Student Support Accelerator looked at AI built into live, chat-based math tutoring. When Google's LearnLM model drafted responses, supervising tutors approved 76.4% of them with little or no editing.21 Students who worked with LearnLM succeeded 66% of the time on later, harder topics. The rate was 61% for students with human tutors alone and 56% for those who got static hints.21 In the second trial, Tutor CoPilot made students four percentage points more likely to master a topic. The gain reached up to nine points for students whose tutors were less experienced or lower-rated.21

The two sets of findings can be reconciled. The successful trials either tightly scripted what the AI did or kept a human tutor in the loop. The disappointing trial simply gave students a general-purpose tutor on top of an existing course. Evidence reviews make a similar point. Gains show up mostly in structured subjects such as algebra and grammar. Very short pilots tend to underperform. And when districts staff both options properly, AI tutors tend to tie or slightly trail small-group human tutoring.29 One review also finds that AI tutors are good at delivering content and practice, but their effect on higher-order thinking is mixed.25

That context is useful because the market is full of bigger claims. Vendor-adjacent roundups cite effect sizes and test-score jumps drawn from case studies, internal reports and quasi-experimental pilots.2230 Some of those numbers may hold up. They aren't comparable to randomized evidence, though. The randomized evidence now includes at least one well-powered study in which simply giving students access made outcomes worse.

K-12 policy swings toward restriction

The biggest school systems are moving toward restriction, not adoption. New York City Public Schools announced a moratorium on student-facing generative AI from pre-K through grade 8 for the 2026–27 school year.8 High schoolers may use only approved, vetted programs, and every high school student must complete two 45-minute AI literacy modules.8 Mayor Zohran Mamdani announced the rules on September 2. Companion chatbots are banned at every grade, and about 5% of high schoolers will take part in five AI pilot programs while the city studies the effects.5 A day later, Los Angeles Unified said it would ban most student AI use on district devices this year.5

Other districts have gone the opposite way. Boston gives students supervised access to approved AI tools starting in third grade and requires disclosure whenever AI is used.5 Denver bans ChatGPT but allows the educational platform MagicSchool.5 Some districts, Alexandria among them, have set up controlled environments: students who try to open ChatGPT are redirected to a filtered chat that refuses to write full essays.9

Statehouses are adding to the patchwork. At least 35 states have issued AI guidance for schools, and Maryland, Idaho, Oklahoma and Virginia passed new laws this year.5 As of March, MultiState was tracking 134 AI-in-education bills in 31 states.9 Ohio and Tennessee now require every district to adopt an AI policy. Oklahoma's SB 1734 requires educator supervision and bars AI from high-stakes decisions.9 Privacy law also limits what schools can do. A FERPA compliance analysis notes that consumer ChatGPT accounts lack the contract terms needed to handle student records, so districts are pushed toward enterprise tiers.4

OpenAI's parallel strategy

While districts restrict student access, OpenAI is expanding into schools through teachers. On August 26, it added 55 school systems in 20 states to ChatGPT for Teachers. The program now covers more than 300,000 educators in over 100 organizations across 30 states, and it stays free through June 2028.10 The expansion came with a shared student-data privacy agreement covering 16 states, which simplifies the compliance work that has slowed districts down.10 The product still doesn't issue student accounts.6 OpenAI also announced a five-year partnership with the American Federation of Teachers to train 400,000 K-12 educators.10

On the student side, OpenAI says that since August 18, users it predicts are under 18 are placed in ChatGPT for Teens by default. That version pushes back on homework shortcuts and steers students toward Study Mode.3 Meanwhile, U.S. college students can claim four free months of ChatGPT Plus through October 31.7 The company is courting teachers and adults while adding guardrails for minors. That positions it to benefit whichever way school policy goes.

Academic integrity: detection gives way to redesign

In higher education, the integrity debate has largely stopped asking whether students cheat with AI. The question now is how to respond without falsely accusing the innocent. Cornell-led research surveyed more than 95,000 students at 20 public research universities. About a third used generative AI regularly for coursework, and 9% had used it to cheat. Among daily users, the rate rose to 26%.16 Co-author Igor Chirikov has said unpublished 2026 data show overall AI use climbing to 80%.17 Other measures suggest more cheating. A Harvard Crimson survey found that a third of graduating seniors admitted using AI against instructor rules. Of those, 93% said they were never caught, and only 2.6% faced any consequence.20

Faculty see the problem clearly but trust the tools very little. In an AAC&U and Elon survey, 78% of faculty said cheating had increased since generative AI arrived. Only 23% rated AI detectors as effective.15 A growing list of universities has stopped relying on detectors. Northwestern, Georgetown and NYU are among those that have switched off Turnitin's AI detection. Yale, Vanderbilt, Johns Hopkins and Indiana discourage using detector scores as the only evidence of cheating.11 Harvard College dean David Deming told students the school would benefit from "getting out of the AI-detection business" because it damages trust.15

Cases on the ground show the costs of both cheating and enforcement. In a 350-student class at UW–Madison, a hidden-prompt trap caught 60 students.15 Some UCLA students reported taking exams on camera with mirrors behind them or their arms crossed. One Los Angeles attorney said AI accusations now make up about 35% of her firm's education caseload.13 Outside the U.S., suspected large-scale cheating forced about 58,000 applicants to retake the entrance exam for Mexico's national university, UNAM.18 In South Korea this month, two universities caught applicants wearing AI-equipped smart glasses during essay exams.19

Institutions are converging on a different approach: grade the process, not just the final product. That means drafts, in-class writing, oral defenses and clear course-level rules about when AI is allowed.914 There is real disagreement over how far to go. UC Berkeley's law school has banned nearly all AI use.13 At Harvard, faculty are split, and science courses are far more likely than humanities courses to allow AI.15

What it adds up to

The research and the integrity crisis are two sides of the same issue. When AI does the thinking, students skip the work that produces learning. The October trial's drop in engagement and the cheating surveys both point to that. When AI is designed to support the work, as in the scripted math-tutoring trials, results improve. Restrictive K-12 policies, process-based assessment in colleges and teen-specific guardrails from vendors are all attempts to keep students doing the cognitive work themselves. The early evidence suggests that whether that works depends far more on course design than on the model.