Sungu’s experiment, conducted with other University of Pennsylvania researchers including educational psychologist Angela Duckworth, followed 193 teachers and more than 2,800 middle and high school students at a chain of private schools in Turkey during the spring of 2025.
Teachers were randomly assigned to receive access to a ChatGPT-based teaching assistant customized to Turkey’s national curriculum or to continue teaching as usual. For 10 weeks, teachers used the tool primarily to generate class notes, assignments, and exams.
Students whose teachers had access to the AI tool rated their classes as less fun, less interesting, and less important than students in the control group. The decline in intrinsic motivation was modest, but larger among students of those teachers who had already been more frequent users of AI before the experiment began.
Average academic performance did not change overall. But among teachers whose students had lower grades before the experiment (an indicator of lower-performing teachers), student performance and confidence declined. Academic performance was measured using externally administered standardized tests, ruling out the possibility that these teachers had different grading standards.
The study cannot explain exactly why the quality of teaching deteriorated. The researchers did not observe classrooms or analyze the AI-generated materials that teachers used. But Sungu suspects that teachers may have been giving up one of their most effective tools.
“When you start using AI-generated material, you’re losing your personal voice,” Sungu said. “Maybe it’s technically good enough, but it doesn’t really convey your own style. If everything is too uniform, it just gets a little more boring.”
One possible explanation for the weaker academic performance among students of low-performing teachers, Sungu said, is that stronger teachers treat the AI results as a first draft, revising them and adapting them to their classrooms. He suspects that weaker teachers are more likely to use AI-generated material as-is.
This study is not a clear comparison between teaching with and without AI. Teachers in the control group were free to use other AI tools, making this a comparison between access to a personalized AI assistant and what teachers decided to do on their own. If anything, Sungu said, these findings could be underestimating the risks of teachers relying heavily on AI-generated materials.
Still, Sungu warns that it would be a mistake to conclude that “AI is terrible and will ruin education.” He sees a different lesson: Access to AI technology alone does not improve teaching.
The challenge is to help teachers use AI in ways that preserve human judgment and creativity. That will require teacher training programs, protective barriers and better interfaces.
“As of now, there’s something to worry about about how teachers are using it organically,” he said.
Sungu says he personally uses AI in his university teaching to create interactive games and surveys that would otherwise take too long to create. “When I first got the result, it looked fantastic,” he said. “And then if I don’t dive into it, the examples, the numbers don’t make sense. I end up spending the same amount of time improving the result or calibrating it for my class.”
“It’s not a time saver,” he said.
This story about AI in teaching was produced by The Hechinger Reportan independent, nonprofit news organization covering education. Enroll in Test points and others Hechinger Newsletters.
