AI Tutoring Study: Kids Barely Use Tools Even With Human Support

The finding nobody set out to test

Pitches for AI tutoring usually focus on what the software can do. A new body of research from Stanford keeps landing on a plainer question: will children actually open it? Across two randomized controlled trials, the answer was mostly no. Adding a human coach didn't change that much.

In an August op-ed for The 74, Meghan McCormick and Carly Robinson summarized the trials led by Stanford's SCALE Initiative. In them, 355 students in grades 1 through 5 across two school districts were assigned either to use an AI literacy tutor on their own or to use the same platform with a human tutor present.1 The researchers describe low usage as the clearest result. Nearly half of the children working independently never used the tutor at all, and those who did logged only two to five minutes a week on average.1 Robinson, who is the study's lead author and the SCALE Initiative's director of research, told Chalkbeat when the paper came out in June that the team hadn't planned to measure this. Having access to the AI tutor, she said, turned out to be different from using it.10

The dosage gap is large. The platform provider said students start to see reading gains after about 30 minutes of use per week, and the study scheduled two 30-minute sessions weekly.22 Among control students in the working paper, only 60.7% in District A and 53.3% in District B ever logged on. Average weekly use was 2.18 and 5.23 minutes.22 Students who did log on used the platform for only four to five weeks of interventions that lasted 14 to 31 weeks.22

What the human tutors did, and didn't, change

The human tutors in this design did not teach reading. They worked with groups of two to five students, helping with motivation, fixing technical problems, setting group norms and leading short check-ins and reflections. Those activities took about half of each session away from screen time.22 In one district the tutors were after-school program staff. In the other they were middle schoolers who were strong readers.10

The intervention did help in a narrow sense. Students with a tutor used the platform about one more minute per week in District A and about 4.4 more in District B. They also completed 71% to 80% more stories than the control groups.22 Those gains came even though tutored students had less time on the device. The authors read this as evidence that relational and organizational support can make up for lost screen time.22 Still, the extra use came to less than two hours over the whole study, and reading achievement did not improve.22 Tutoring also did not make students any more likely to log on in the first place.10

The coverage divides on how to frame this. District Administration's headline said human support "changed" students' non-use of a free AI tutor, and it emphasized that human involvement significantly raised engagement.6 The authors' own op-ed is blunter. Its subhead says human support didn't get kids engaged, and the piece says the help fell well short of recommended usage.1 Both readings are technically accurate. The second one matters more for schools. A 71% to 80% increase on a very small base is still a small number, and a district paying for a platform is paying for learning, not for relative gains in stories read. Robinson told Chalkbeat that the study never got close enough to the needed dosage to say whether the AI tutor works.10 That is the most honest summary of the result.

A pattern, not an outlier

One trial would be easy to dismiss. This one fits a pattern. Chalkbeat reported a separate two-year experiment in 18 Tennessee middle schools. Low-performing students who used Khan Academy turned to its AI tutor, Khanmigo, on only about a third of the days they spent on the platform.8 The researchers wrote that access was nearly universal but engagement was thin.8 Sal Khan, who had earlier called the AI tutor a "non-event" for most students, said the results showed how much motivation matters.8

Government Technology described Brown University's analysis of the same Tennessee experiment. Students on Khan Academy improved in math, but researchers found little evidence that the AI tutor caused the gains. They credited more structured, individually targeted practice instead.5 A Stanford brief from August adds a larger data point. In a study of 181,000 students on a supplemental math platform, only 5% reached the recommended 30 minutes a week and 41% never logged on. Teacher, school and district factors explained 57% of the variation in usage.

One result points the other way. A 2025 randomized trial published in Scientific Reports found that college students learned more in less time with a carefully designed AI tutor than in an active-learning class, and said they felt more engaged and motivated.21 The difference is probably the population and the setting. Undergraduates in a structured course session are not first-graders who are expected to log on by themselves after school. The Stanford work is about K-12 students who have to start on their own, and that is where the engagement problem shows up.

The equity problem inside the usage gap

The most troubling detail is who used the tools. In the Stanford trials, control students who used the AI tutor were more likely to be higher achieving and less likely to receive special education services. The struggling readers who might have gained the most were the least likely to engage.22 The authors note this could partly reflect accommodation decisions for individual students. Even so, they warn that if AI tutors are mostly used by students who are already doing well, the technology alone may not close gaps.22

Chayne Turano of SCALE made the same point to Government Technology. Students who started using the tools tended to be the ones already succeeding in class, which makes an existing equity problem worse.5 Cost is part of the tension. The Brown study put Khanmigo at about $15 per student per year. High-dosage human tutoring can cost several thousand dollars per student but produces larger gains.5 Stanford's Ana Trindade Ribeiro warned against treating those numbers as a case for replacing human tutors.5

Where academic integrity enters the picture

The Tennessee data also show how students use these tools when they do engage. Many of their messages to Khanmigo were off-topic, and many students tried to get the chatbot to give them the right answers.8 Futurism listed examples such as asking for an essay on the history of pizza or asking for a joke.7 Researcher Philip Oreopoulos said the technology is ready for an engaged student who wants a good tutor, but most students don't use it that way.8

This matters for the integrity debate in a less obvious way. The usual worry is that AI will set off a wave of cheating. Stanford-affiliated survey research on high schoolers found self-reported cheating holding at 72.06% a year and a half after ChatGPT's release, which is in line with historical levels.13 Among undergraduates, a Berkeley-led survey of more than 95,000 students found that at least 9% of AI users said they had used it to cheat. The share rose to 26% among daily users, compared with 7% among monthly users, though the researchers could not say which caused which.14 In schools, Harvard researchers report that suspected AI-assisted work accounts for most of the AI "dilemmas" teachers and principals describe, and that this strains relationships between students and teachers.17

Taken together, the tutoring and integrity research describe the same student behavior. Students look for the easiest route. They skip a tutor that asks them to work, and they try to pull answers out of one that is supposed to withhold them. Chalkbeat observed that adolescents often take the path of least resistance, and most of their time is spent outside tightly controlled settings.8 Designing a tutor that "won't give away the answer" doesn't help much if students either ignore it or treat it as a game to beat.

What might actually work

The more promising evidence comes from AI built for adults. In one large randomized trial, students whose human tutors had an AI tool offering real-time suggestions were 4 percentage points more likely to master lesson topics. The gain rose to 9 points among students of lower-rated tutors.5 Stanford's brief states that the strongest results so far come from AI tools designed for tutors rather than students.7 A smaller study by Oreopoulos and colleagues found some benefit when an AI helper appeared without being asked and students had to answer correctly before moving on. Those students were slightly more accurate on a quiz a week later.8

The evidence base is still thin. A SCALE review found more than 800 studies of AI in education, and only 20 used designs rigorous enough to estimate causal effects.2 McCormick and Robinson say Overdeck Family Foundation grantees, including OnYourMark and the University of Chicago Education Lab, are testing hybrid models that combine AI personalization with targeted human support. Whether those models work at scale is still unknown.1

The research so far supports one conclusion: student engagement is the bottleneck, not the quality of the model. As several states run AI tutoring pilots3, districts should judge a product less by whether it can tutor and more by whether their own students will actually use it, and who will make sure they do. Until vendors can answer that, the safer spending is on AI that helps the adults who are already working with students.