AI Tutoring Research Splits as New Trial Finds Lower Grades

The verdict that set the terms

This year's most influential warning about artificial intelligence in schools did not come from a single classroom experiment. It came from a large qualitative study. The Brookings Institution's Center for Universal Education published A new direction for students in an AI world: Prosper, prepare, protect and concluded that, for now, the risks of generative AI in children's education outweigh its benefits.7 The authors call the work a "premortem." It drew on interviews, focus groups and consultations with more than 500 students, teachers, parents, education leaders and technologists in 50 countries, plus a review of more than 400 studies and a Delphi expert panel.7

The report's central argument is about the kind of harm involved, not only the amount. In its framing, the benefits of AI tend to be additions to learning, while the risks can damage the conditions that learning depends on.5 The authors say AI threatens children's foundational development, and that this can stop the benefits from ever showing up.7 Coverage of the findings reported that feedback from participants leaned negative overall: 56 percent of responses pointed to harms and 44 percent to benefits.5

Nine months later, a randomized trial at a large U.S. public university has tested this concern directly. Its results are harder to dismiss than the opinions of focus-group participants.

What the report says goes wrong

Reporting on the Brookings study across outlets settled on a few main harms. The first is what the authors call cognitive offloading. Forbes described their phrase for it, a "great unwiring" of students' cognitive abilities, and their claim that AI has "turbocharged" the habit of handing thinking over to a machine.4 NPR reported that the review found evidence of declines in content knowledge, critical thinking and creativity among students who use generative AI. One student summed up the appeal by saying you don't need to use your brain.1

Rebecca Winthrop, who leads the Brookings center, has tied the problem to what she calls "wild west" use: students talking to chatbots outside school, without supervision and with content nobody has vetted.2 She has also pointed to companion bots built to be sycophantic. Her worry is that children who are constantly flattered may struggle to accept feedback, which she sees as central to teaching.10

The second harm is to trust, and The 74 flagged it as one of the report's most striking findings. Teachers doubt that students' work is their own. Students, in turn, suspect teachers who use AI to build assignments without saying so.8 In interviews, 16 percent of teachers called this loss of trust a significant concern.8

The third is equity, and here the report argues something new. Winthrop has said this may be the first time in ed-tech history that schools have to pay more to get accurate information, because premium models are more reliable than free ones.18

The report is not anti-technology. It credits AI with helping second-language learners, students with dyslexia and other disabilities, and girls in Afghanistan who are barred from school and receive curriculum lessons over WhatsApp.1 Winthrop told Education Week that the real benefits appear when AI is "held back," used narrowly and paired with vetted content and sound teaching.10

Where the coverage diverges

News outlets mostly passed along the report's conclusions without challenge. The 74, however, pointed out an admission inside the report: researchers have little rigorous, long-term evidence on how AI affects learning and wellbeing, and the authors concede that nobody can predict its effects with full accuracy.8 A reader comment on the Forbes column was sharper, arguing that the conclusions rest on the subjective views of people already in the system rather than on outcome data.4 Forbes columnist Peter Greene questioned the report from the other side. He argued that some of the "benefits" it endorses, such as AI-written letters to parents or AI help with editing drafts, would trouble many educators.4

That criticism matters because experimental research on AI tutoring has often pointed the other way. A Harvard trial in undergraduate physics, published in Scientific Reports in 2025, found that students using a custom AI tutor learned substantially more in less time than students in an active-learning class. Coverage reported effect sizes between 0.73 and 1.3 standard deviations.1618 Stanford's National Student Support Accelerator described two recent randomized trials of AI built into live, chat-based math tutoring. In one, supervising tutors approved 76.4 percent of Google LearnLM's responses with little or no editing. Students tutored with LearnLM's help reached a 66 percent success rate on later, harder topics, compared with 61 percent for students tutored by humans alone.12 In the other trial, Tutor CoPilot made students four percentage points more likely to master a topic. The gains reached nine points for students whose tutors were less experienced.12

Older research on intelligent tutoring systems is also encouraging. A widely cited meta-analysis found these systems beat large-group, teacher-led instruction (g = .42) and performed about as well as one-on-one human tutoring.14

On the surface, then, there are two camps. One is a global consultation warning of cognitive decline. The other is a set of controlled trials showing real gains.

The trial that bridges the gap

An EdWorkingPaper released this month weakens that split. It reports what its authors describe as the first large-scale randomized experiment on AI tutoring at a big U.S. public university: 2,379 undergraduates and 30 instructors across several disciplines in fall 2025.17 Students were given access to a generative AI tutor built into their courses. Comparing sections of the same course, that access lowered final grades by 0.37 standard deviations and cut participation in instructor-designed online activities by 0.90 standard deviations.17 First-generation students lost more than twice as much in final grades as their continuing-generation classmates.17

Some details call for caution. Only about 15 percent of students offered the tutor used it, averaging about four sessions. The authors say this raises questions about what caused the drop, including whether simply being offered the tool changed how students studied.17 In the full sample, which includes courses with poorly matched grading, the evidence for harm was not significant. The clearest effects showed up in the exact-match sample.17

The paper's literature review is more useful than any single number in it. It describes a growing pattern called the "learning penalty": general-purpose AI helps students on assisted tasks but hurts their later performance when working alone. Purpose-built tutoring guardrails can reduce that harm.17 The authors set the Harvard physics result next to studies finding that unrestricted ChatGPT use encourages offloading and weakens retention. Their conclusion is that AI's effects depend on how the tutor is configured, the instructional scaffolding around it, and how students actually engage with it.17 They also cite work suggesting that Khanmigo produced modest gains but little sustained use, and that an AI-enhanced practice platform led to slower progress than the same platform without AI.17

The reading: design, not technology, decides

Taken together, the evidence supports one conclusion. The headline question of whether AI helps or hurts learning has no single answer, and the Brookings report and the positive trials are measuring different things. The trials with strong results tested narrow, carefully designed tutors in structured subjects, often with a human involved. Those are the "held back" uses Winthrop herself supports.1012 The report's warnings are aimed at open-ended, unsupervised chatbot use, which is what most students actually encounter.2 Analysts tracking tutoring trials this year likewise find that gains cluster in structured areas like algebra and grammar. In those reviews, AI tutors tie with or slightly trail well-staffed small-group human tutoring.13

The university trial matters because it closes the gap between the two kinds of evidence. A tool presented as a tutor and placed inside real courses still seems to have pulled students away from the activities instructors designed, with no academic payoff.17 In other words, calling something a tutor does not make it work like one.

The practical stakes are large, because AI use is already nearly universal. Education Week, citing the Center for Democracy and Technology, reported that 85 percent of teachers and 86 percent of students used AI in some form during the 2024-25 school year.10 Brookings recommends that districts agree on shared procurement criteria, on the theory that vendors will follow what buyers demand.2 It also calls for AI built for young people to be less sycophantic and more willing to push back.1 The new trial evidence makes that procurement idea more pressing. If a tutor's design determines whether it builds students' skills or replaces their effort, schools need proof of learning outcomes before they buy, not after it has been rolled out to every classroom.