Expert system is often promoted as a way to make instructors more reliable by helping them write lesson plans, create class products and supply feedback to students in seconds. But among the very first randomized trials testing AI in genuine class discovered that it can likewise weaken knowing. Students whose instructors were given access to an AI mentor assistant feltless motivated to find out.
The damage was particularly noticable amongst students whose teachers were already weaker trainers, as determined by their performance before the experiment started. Their trainees also scored lower on standardized final examinations, the scientists found.
“Teachers, similar to trainees or coders, may be using AI as a crutch,” stated Alp Sungu, lead author of the research study and an assistant professor at the Wharton School at the University of Pennsylvania. “Instead of doing the actual work, they’re utilizing AI to hand over the task, and that lowers the quality of their teaching.”
A draft of the research study, “Generative AI Can Damage Mentor,” was released online in June and has not yet been published in a peer-reviewed journal. It echoes Sungu’s extensively discussed 2024 research study on how trainees’ use of AI is damaging knowing.
“Students utilize AI as an answer device, not as a tool for knowing, and therefore it hurts knowing,” stated Sungu. “Here, I think teachers are potentially utilizing AI as a product producing maker for research, lecture notes, lesson strategies, syllabus. Instead of improving their own output, they’re using AI as a replacement with extremely minimal interaction, and therefore the quality of output is not good enough.”
Related: Faster services, lower test scores: How AI is eroding mathematics skills
Sungu’s experiment, performed with fellow University of Pennsylvania researchers, including educational psychologist Angela Duckworth, followed 193 instructors and more than 2,800 middle and high school trainees in a private school chain in Turkey during the spring of 2025.
Teachers were arbitrarily designated either to get access to a ChatGPT-based mentor assistant customized to Turkey’s national curriculum or to continue teaching as typical. Over 10 weeks, instructors primarily used the tool to generate lecture notes, projects and exams.
Trainees whose instructors had access to the AI tool ranked their classes as less pleasurable, less interesting and lesser than students in the control group. The decline in intrinsic inspiration was modest, however bigger among trainees of those instructors who had actually currently been much heavier AI users before the experiment began.
Average scholastic achievement did not alter total. But amongst instructors whose trainees had lower marks before the experiment– a proxy for lower-performing teachers– trainee accomplishment and confidence both declined. Academic achievement was measured through externally administered standardized tests, eliminating the possibility that these instructors had various grading requirements.
The study can not explain precisely why mentor quality degraded. Researchers did not observe class or evaluate the AI-generated products instructors used. But Sungu believes that instructors may have been quiting among their most efficient tools.
“When you start using AI-generated material, you’re losing your personal voice,” stated Sungu. “It might be technically good enough, however it doesn’t actually carry your own style. If whatever is very consistent, it simply becomes a bit more dull.”
One possible description for the weaker scholastic efficiency among trainees of low-performing instructors, Sungu said, is that more powerful teachers deal with AI output as an initial draft, modifying and adapting it to their classrooms. Weaker instructors, he suspects, may be most likely to use AI-generated material as is.
Related: AI gives more appreciation, less criticism to Black students
This research study is not a clean comparison between mentor with and without AI. Educators in the control group were free to utilize other AI tools, making this a contrast in between access to a tailored AI assistant and whatever teachers chose to do on their own. If anything, Sungu stated, these findings may be downplaying the risks of teachers relying heavily on AI-generated products.
Still, Sungu warns that it would be a mistaketo conclude that “AI is terrible and will ruin education.” He sees a various lesson: Access to AI innovation alone does not improve mentor.
The difficulty is to assist instructors use AI in ways that protect human judgment and creativity. That will require instructor training programs, guardrails and much better interfaces.
“As of today, how teachers are utilizing it organically, there is something to be stressed over,” he said.
Sungu says he personally utilizes AI in his university teaching to develop interactive games and polls that would otherwise take too long to construct. “When I initially get the output, it just looks fantastic,” he stated. “And after that, if I do not immerse myself in it, the examples, the numbers do not make sense. I wind up spending an equivalent amount of time to improve the output or adjust it to my class.”
“It’s not a time saver,” he stated.
Contact personnelauthor Jill Barshay at 212-678-3595, jillbarshay.35 on Signal, or [email protected].
This story about AI in teaching was produced by The Hechinger Report, a not-for-profit, independent news organization that covers education. Register for Proof Pointsand other Hechinger newsletters.
Was this story helpful? Leave a suggestion to support your education reporters.
The Hechinger Report is a nonprofit newsroom powered by reader support
Republish our posts for free, online or in print, under an Imaginative Commons license.