Advocates for the use of AI in the classroom have long argued that the tech could supercharge learning, giving the current generation of students a major leg up as they get ready to enter a nearly unrecognizable work force.

But reality has long struggled to keep up with those lofty visions. Educators have watched in horror as students in the AI era struggle to read long documents, complete their homework without the use of a chatbot, or pass in-person exams. Meanwhile, early claims that tools like OpenAI’s ChatGPT could have positive impact on learning performance are being retracted wholesale.

In an intriguing new wrinkle for the debate, young students appear to be rejecting AI tutors outright. Scientists spent two years offering use of the Khan Academy’s Khanmigo generative AI tutor to students across 18 Tennessee middle schools, as detailed in a new paper. The tool was meant to be used during the school day, and was directed not to give up correct answers in order to slowly guide students through the learning process.

The results probably won’t be shocking to anyone familiar with rebellious 12-year-olds. In practice, the students mostly ignored the tool — and when they did use it, it wasn’t for anything to do with their schooling. Instead of using Khanmigo to seek help during their online lessons, most students sent the chatbot off-topic messages, messing with it to test its limits. Many also tried to coerce it into giving up correct answers.

Students peppered the chatbot with unrelated prompts, like “can you write an essay about the history of pizza?” or asking it to “say a joke.”

“Access was nearly universal but engagement was thin,” the researchers wrote. “It may be that human attention is the key ingredient for realizing the full benefits of personalized learning.”

The findings once again highlight the importance of human tutoring in the age of AI as companies and school systems continue to shoehorn the tech into textbooks, curriculums, and other aspects of student life.

“If a keen, engaged student wants to use the technology as a good tutor, it’s ready to go,” coauthor and University of Toronto researcher Philip Oreopoulos told education publication Chalkbeat. “The issue is that most students don’t use it that way.”

“What this has put in huge spotlight — not that it’s a shock to educators — is how much motivation matters,” Khan Academy CEO Sal Khan told Chalkbeat.

The researchers did find that students using Khanmigo progressed faster in math when compared to a control group that didn’t use the AI. However, whether that’s enough to justify the use of AI chatbots in the classroom is debatable at best.

The results of the study echo other findings. Earlier this month, researchers at Stanford concluded that AI wasn’t a fix-all. Their takeaway: the tech won’t succeed without human guidance.

“The effectiveness of AI-led tutoring depends as much on integration, such as by teachers in classrooms or parents at home, as on the quality of the software itself,” the researchers found.

“The strongest results so far come from AI tools built for the human tutors, not the students,” a summary of the Stanford-led research reads.

It’s a precarious balance between giving students enough space to engage with their learning material while providing enough of a push via human educators. An AI chatbot alone isn’t going to succeed at teasing out that balance any time soon, the evidence suggests.

“The same technology that can be used as a good tutor, can also be used to make your life easier,” Oreopoulos told Chalkbeat. “But making your life easier equals less effort, less learning.”

More on AI and education: The Results of Middle and High School Students Using AI Are Extremely Ominous