For two years, students in a randomized study had an artificial intelligence (AI) tutor one click away. Of those students, 96% tried it at least once. However, the median student only ended up using the tutor, called Khanmigo, a third of the days they had access to it.

That is the core finding of “One Click Away,” a National Bureau of Economic Research working paper by economists Philip Oreopoulos and Nina Low that followed Khan Academy’s Khanmigo in 18 Tennessee middle schools. Students used Khan Academy during daily math intervention sessions, with Khanmigo set up to coach them rather than hand over answers.

Compared with students in the usual remedial instruction, the students in the intervention sessions made modest math gains, about what you would expect from Khan Academy practice without AI. Most students used the AI tutor infrequently, and their conversations with it were rarely about math. In other words, they made little effort to benefit from it.

In a blog post, Khan Academy CEO Sal Khan walked through the study and noted that low use of Khanmigo matched what his organization had been saying publicly. In a recent interview with Chalkbeat, he said the study showed “how much motivation matters.”

Teachers won’t be surprised by that, though a lot of money has been spent on the opposite assumption: that if you build a patient, capable AI tutor and put it in front of students, they will use it. Microsoft donated enough Azure infrastructure to make Khanmigo’s teacher tools free for every educator in the country. And states bought licenses in bulk — for example, Maryland bought 4,350 of them for a two-year pilot and Indiana launched a program across 112 schools.

How Khan Academy Adjusted

Khan Academy has already changed course. In a July blog post, Khan wrote that the AI tutor couldn’t just sit next to the content; it had to be woven into it, and the team had to “make productive struggle harder to sidestep.” Earlier this year, the organization built Khanmigo directly into its practice problems rather than waiting for students to ask for it.

That fix makes sense for Khan Academy, but the underlying point is easy to misread as “every student needs an AI tutor.” They don’t. Students can learn calculus, write strong essays and think well without any AI at all, and plenty of teachers will reasonably keep it out of their classrooms.

If a school wants AI to support student thinking, use of AI can’t be voluntary. When AI help is optional, students don’t choose the tutor that asks them questions. They choose the chatbot in the next tab that gives them the answer. I’ve seen that pattern in my own work for two years, and the Tennessee study shows it at scale. When I have delivered AI study tools to students as an option in college and high school classes, they have consistently deferred engagement. But when I required students to use AI tools and told them I’d be monitoring the conversations and giving them feedback, I saw the beginnings of meaningful engagement.

Even then, requiring use of the tool only solves the participation problem, and not the effort problem. A student can be required to talk to a tutor and still give it nothing. Khan Academy’s chief learning officer, Kristen DiCerbo, said on an EdTech Insiders panel in February that she was seeing “more passive kinds of interaction than we would like.” A built-in tutor will start more of those conversations. It can’t make the student care how they do.  

Requirement vs. Assessment

I think the field is missing a step: treating the conversation itself as the assessment.

That was the design of the research in “From Tutor to Test,” my chapter with Aimée Skidmore in the Elsevier book “GenAI in Higher Education.” The study followed 21 students in 12th grade in a Media and Communications course at Collège du Léman International School in Switzerland. Over four weeks, students compared transcripts of strong and weak interaction with AI side by side, then worked with their teacher to build a rubric for what effective AI interaction looks like. They then used ChatGPT to brainstorm for a media analysis project and turned in their full transcripts, which were evaluated using that rubric. The students were graded on how precisely they asked questions, how critically they responded to the AI’s suggestions and whether they kept ownership of the thinking.

Afterward, 18 of the 21 students said their approach to AI had changed, and 10 said they now think more about their own input and how to build on it later in the conversation. The number who felt completely confident getting good responses from the AI rose from two to 11.

Rising confidence can be a warning sign, so we tracked students individually. Five gained major confidence while also becoming more aware of AI’s risks. Four became less confident in their ability to judge the AI’s accuracy, which we read as healthy recalibration.

Two showed the pattern we worried about: big confidence gains paired with a dismissive attitude. One student, who had never used AI before, wrote that students “need to know how to approach AI to not lose their capacity to think.”  

It was a small pilot without a control group, and most of the data is self-reported, so I treat it as a starting point. Still, the logic is straightforward. When the conversation is the test, students know what they have to show and are motivated to make an effort because someone will judge the results.

Firsthand Experience

Six months after Skidmore and I ran that study, I turned the experiment around on myself. As part of my master of fine arts research on AI and creativity at Wilkes University, I used AI in my own coursework in two ways: as a conventional Socratic tutor, and as character role-plays in which the AI portrayed a specific, difficult person I had to persuade, teach or negotiate with. I asked my professor to evaluate my conversations and discuss them with me. The goal was to see whether either approach deepened my own critical thinking, writing skill and creativity.

The tutor bored me, and I understood right away why students avoid it. Given the choice, I would rather have done the work on my own, or taken a shortcut through a more compliant chatbot.

The role-plays were more engaging, but that wasn’t the most important difference. I knew my professor was going to evaluate my performance in the chat, which made me want to do well, and doing well meant thinking harder than I otherwise would have. Said differently, the knowledge that I was being monitored raised the stakes of the engagement. The AI was the same in both versions. What changed was my awareness that someone would judge my work.  

Classroom Examples

Based on both sets of research, I built a platform to continue testing the pedagogical value of the approach. I called it AI Friction Labs, and the role-plays are called Friction Bots.

Full disclosure: AI Friction Labs is my company, and the classrooms below were beta testers. What follows is early product feedback, not research.

In Bill Fallon’s calculus class at Cheshire Academy in Connecticut, 21 students ran a role-play in which they played the teacher. The AI played “Devon,” a classmate who didn’t understand tangent lines and wouldn’t pretend to. When students leaned on terms such as “instantaneous rate of change,” Devon got more confused. To succeed, students had to understand the concept well enough to explain it in plain language. The transcripts were then evaluated — first by an AI engine, then by the teacher — for student persistence, clarity and whether the student stayed in the teaching role.

Five students chose to retake the role-play simulation. All five scored higher the second time, by an average of about 34 points. It’s a small group, but the reason they went back matters: the transcript and the AI evaluation showed them where their explanation had broken down.  

Other teachers have seen similar results. In Mary Churchill’s science classes at San Domenico School in California, an AI playing the role of a seasoned ecologist kept pushing students with “why?” and “how?” follow-ups. One student said it “forced me to go beyond memorizing terms and actually understand the mechanisms behind them.”

Not Everyone Likes It

Some students disliked these exercises. In Churchill’s classes, one felt “interrogated.” Another said the questioning was close to “rage bait.” Several said it got repetitive. Some of those complaints point to real work: difficulty that needs tuning, better scaffolding and a teacher watching for the student who is overwhelmed rather than challenged.

Some of that frustration is expected. Fallon saw the same split in his class. “Things are going to frustrate you in life,” he says. “People are going to frustrate you. This bot is frustrating you. How do you work around that and make it so that you are productive and helping the person as best you can? That is a whole other level of thought and learning.”  

Students don’t love tests either. We don’t usually take that as evidence that tests measure nothing.

Where This Leaves Us

Khan is right that motivation matters. Schools can decide AI has no place in a class, and that’s a legitimate choice. But where teachers do want AI to support thinking, making it optional doesn’t work, because the option students take is the one that does the thinking for them. Requiring use of AI isn’t enough on its own either. Students work hard when they know what they’re supposed to show and that someone whose judgment they respect will look at it. That was true of good assessment long before AI.

Much of the original AI tutor vision rested on self-directed learning: a student, motivated from within, alone with a patient machine. The model I’m describing puts the teacher back at the center. The system produces a chat transcript, which shows how a student reasoned in real time, where they got stuck and how they recovered. When the transcripts come in, AI can draft a first-pass evaluation against the criteria, but the teacher makes the call. They add notes, correct the draft and decide what goes back to the student.  

After having graded thousands of pages of student chats manually over the past three years, I see those tools are more than just mechanisms to keep me at the center of instruction; they also assist in the sometimes-burdensome task of sifting through piles of transcribed conversations. The AI engine evaluates transcripts against specified criteria, and then highlights specific messages where the student may have succeeded or struggled in achieving the goal of the role-play. Those “teachable moments” help me to zero in on key moments in the role-play and evaluate them against the AI feedback. 

Further, as a test in and of itself, the role-plays are not run daily or even weekly, but rather periodically and with meaningful preparation and practice beforehand. 

Schools have asked how to stop students from using AI to avoid thinking. A better question is how to design work where their thinking is visible. No design is cheat-proof; a determined student could feed a character’s lines into another chatbot. But when the assessment is a live, adaptive conversation judged on how the student reasons, there is much less to outsource, and the transcript makes it easier to tell whose thinking it was.

AI makes it possible to build assessments that resemble real situations: a confused classmate, a skeptical reviewer or a colleague who won’t sign off. The student has to use the target skill to get through the conversation, and the teacher gets a record of how it went. That’s a better use of the technology than a tutor waiting at the edge of the screen, whether students click on it or not.