Most of the debate about artificial intelligence (AI) in schools is about what pupils do with it. A newer question is what happens to pupils when their teachers lean on AI-generated lessons. One of the largest trials so far, run by researchers at the University of Pennsylvania in 14 schools in Turkey, found that pupils whose teachers were given an AI planning assistant rated their classes as less enjoyable, less interesting and less important. Nobody told the pupils their teachers were using it.
The short version: giving teachers a ChatGPT-based assistant for about ten weeks did not change average exam results, but it nudged pupils’ motivation down. For pupils of teachers who were already weaker, exam scores and confidence fell as well. Most teachers used the tool to produce finished materials, with a median of two prompts per conversation. The lesson for schools is about how teachers plan with AI, not whether they do.

What the trial tested
The working paper, Generative AI Can Harm Teaching, is by Alp Sungu, Benjamin Lira and Angela Duckworth. It is dated 25 June 2026, sits on the Social Science Research Network (SSRN) and has not yet been through peer review. The full paper is also included in the authors’ open replication package on Zenodo. It is back in the news after Edutopia covered it on 19 November, and Stanford University’s AI Hub for Education included it in its November 2026 review of causal evidence on AI in schools.
The researchers worked with a private school network in Turkey. In January 2025, between the autumn and spring terms, teachers were randomly assigned, department by department, to one of three groups: business as usual, access to a custom AI teaching assistant with a one-hour training session, or the same access plus weekly emails showing how much they had used it. The assistant was a chatbot linked to curriculum resources from the Turkish Ministry of Education. Teachers chose their subject and year group and could ask for help with lesson planning, assessments, feedback, differentiation and admin.
The analysis covers 193 teachers and 2,816 pupils in grades 5 to 12 (roughly Year 6 to Year 13 in England) across 14 schools and 11 subjects. Achievement came from centrally set exams, and motivation and confidence from an end-of-trial survey of 2,343 pupils.
What AI-generated lessons did to pupil motivation
Pupils of teachers with access to the tool reported lower intrinsic motivation, by 0.11 standard deviations. That is modest. Still, all five survey items moved in the same direction, and the largest drop was in how important pupils thought the subject was. Confidence dipped too, though that result was only marginally significant.
Two details make it worth a teacher’s attention. The fall in motivation was larger where teachers were already heavy AI users before the trial, an estimated 0.18 standard deviations against 0.05 for lighter users. And while average exam results did not change, that average hid a split. For pupils of teachers whose classes had scored below the department median the previous term, exam scores fell by 0.13 standard deviations and confidence fell as well. Pupils of stronger teachers showed no significant change.
Pupils may not be able to spot a chatbot in your worksheet, but they seem to notice when the lesson stops sounding like you.
How the teachers actually used the tool
Two thirds of conversations (66%) were about producing teaching materials: lecture notes and lesson content made up 32% of all conversations, homework and exams 22%, and syllabus design 9%. Only 16% were about instructional support such as differentiation, correcting misconceptions or giving feedback. The median conversation had two prompts.
So most teachers asked for a finished product and took it. Sungu told The Hechinger Report that teachers may be using AI “as a material generating machine” with “very minimal interaction”.
Picture a Sunday evening. You ask for a Year 8 lesson on photosynthesis and get a tidy starter, a worksheet and an exit ticket and paste them into your slides. Nothing is wrong. But the lesson no longer has the story about the dying basil plant on your kitchen windowsill, the misconception your last class held on to, or the example that only works with 8C. The authors suggest exactly this: AI materials can be technically adequate yet lack the teacher’s personal voice, and when many teachers use the same tool their materials may start to converge.
Why weaker teaching suffered most
The authors’ second explanation is skill substitution. Teachers who are less secure in their subject or their planning may be the least able to spot what is weak in an AI draft, and the least likely to rehearse it and make it their own. It is the same cognitive offloading worry we usually raise about pupils, turned back on us. Sungu suspects stronger teachers treat the output as a first draft while weaker teachers use it as it comes. The data fit that idea, but nobody observed the lessons or analysed the materials, so it remains a hunch.
Set this beside other trials and a pattern starts to appear. Stanford’s review describes a 2024 trial in England with 259 Year 7 and 8 science teachers in 68 schools, where ChatGPT plus a guide cut lesson preparation by about 25 minutes a week with no drop in quality. I looked at similar time-saving claims in the Aila lesson planning trial. In the United States, Tutor CoPilot, which gave live AI suggestions to 900 tutors working with 1,800 pupils, helped the pupils of lower-rated and less experienced tutors most. My reading is that a tool which coaches the adult in the moment behaves very differently from an open-ended generator that hands over finished lessons.
The caveats worth knowing
This is one study, and it has real limits:
- It ran for a single term, with one tool, in one private school network in one country.
- Control teachers were free to use other AI tools, so it compares access to this assistant with whatever teachers did anyway.
- Exams sat close to the ceiling, with a control group average of 89.2 out of 100, which makes gains hard to detect.
- There was no pupil survey before the trial, so motivation was only measured at the end.
- It is a working paper that has not yet been peer reviewed.
None of that makes the findings go away. It does mean the honest headline is “AI access alone did not improve teaching and may have harmed some of it”, not “AI ruins lessons”.
How to plan with AI without losing your voice
Write the parts pupils hear yourself
Your hook, your core explanation, your worked examples and your stories are where pupils meet you. Keep those yours. AI is far safer on the parts pupils experience as admin: extra practice questions at three levels, a reformatted reading list, a first pass at a mark scheme you will then fix.
Go past two prompts
Argue with the draft. Ask what misconceptions it has missed, tell it what your class found hard last time, and rewrite the explanation in your own words. If you would not be comfortable teaching it without notes, it is not ready.
Run a voice check before you teach
Before the lesson, ask one question: what in this could only have come from me? If the answer is nothing, add something. A local example, a question from last year’s class, a story about your own mistakes.
What school leaders should take from it
The trial gave teachers a tool and a one-hour session. Many schools do much the same, and it was not enough. Education Week reported this month that about two thirds of the 852 educators in its survey had attended at least one professional development session on AI, up from 56% a year earlier, and teachers it interviewed still wanted more hands-on practice. Training should model drafting, questioning and adapting, rather than showing how fast a tool can produce a worksheet.
Three practical moves follow. Look at how staff use AI as well as whether they use it. Give early career teachers and anyone struggling with planning a mentor who reviews AI-assisted lessons with them, because they are the group this study suggests is most at risk. And ask pupils: a few pupil voice questions on interest and importance, before and after any rollout, would tell you more than usage dashboards. If the aim is to cut teacher workload, be honest that the saving may be smaller than promised. Sungu told Hechinger that for his own teaching, once he fixes the AI output, “It’s not a time saver.”
Frequently asked questions
Do lessons generated by artificial intelligence (AI) harm pupils’ learning?
In this trial, average exam results did not change. Pupils’ motivation fell slightly, and pupils of lower-performing teachers did worse in exams and felt less confident. The study is an unreviewed working paper from one country, so treat it as a serious warning rather than a final verdict.
Did pupils know their teachers were using AI?
No. Pupils were not told that their teachers had access to the AI assistant. The authors suggest pupils may have picked up on materials that felt more generic and less like their usual teacher, but the study did not test that directly.
Should teachers stop using AI for lesson planning?
The evidence does not say that. An earlier trial in England found that ChatGPT with a guide saved science teachers about 25 minutes a week with no drop in quality. The risk seems to come from accepting finished materials with little editing, so use AI for drafts and keep your own explanations and examples.
What AI tool did the teachers in the study use?
Teachers used a ChatGPT-based chatbot linked to curriculum resources from the Turkish Ministry of Education. It had options for lesson planning, assessments, feedback, differentiation and admin. Teachers received a one-hour training session, and one group also got weekly emails about their usage.
Have AI-generated lessons changed how your classes feel, to your pupils or to you? Tell me what you have noticed in the comments.



