Large-scale research shows how AI goes wrong in the classroom without didactics and AI literacy

A large-scale field experiment in secondary schools in Turkey shows that a generic AI assistant does not automatically improve lessons. Pupils found the lessons less enjoyable, and with less didactically skilled teachers, results even declined. For us, this research mainly confirms that AI in education only works in combination with strong didactics and genuine AI literacy. That is precisely what EduGPT is built upon.

What was investigated

Researchers from, among others, Wharton and the University of Pennsylvania followed 193 teachers and over 2,800 pupils at fourteen schools. The teachers were randomly divided into two groups.

One group received an AI assistant based on GPT-4o, filled with the national curriculum, plus one hour of training. The other group prepared their lessons as usual.

The outcomes

Pupils of teachers with the AI assistant rated their lessons as less enjoyable, less interesting, and less important to perform well in. Their intrinsic motivation was on average 0.11 standard deviations lower than in the control group.

There was no average difference in exam results. However, among the less didactically skilled half of the teachers, pupils with AI support actually scored lower, and their self-confidence also declined.

It was also striking how the assistant was used. Two thirds of the conversations were about producing lesson materials, and only a small part about didactics, even though the tool also had features for differentiation and feedback. The average conversation consisted of a single request for a lesson plan and at most a follow-up question, after which the outcome was taken into the classroom almost unchanged.

What was actually measured here

This research does not measure what AI can do in education. It measures what happens when a generic chatbot without didactic design ends up with teachers who are barely prepared for it.

After an hour of training, you are not AI literate. Everywhere there are calls to critically assess AI output, but in this experiment almost everything was adopted without scrutiny. And the group where things went most wrong also had the least didactic expertise.

Good teaching in an age of AI requires a combination of didactics, pedagogy, and AI literacy. If any of these three are lacking, an AI assistant yields little or may even be counterproductive.

How EduGPT does things differently

This gap is precisely the reason we built EduGPT. EduGPT is not a bare language model with a curriculum inside, but an environment in which teachers themselves design AI companions in which their didactic choices are embedded, from teaching methods to the way feedback is given.

A well-designed companion asks questions before giving answers. What has already been covered, what can this class already do, where are pupils struggling? This way, the teacher remains the designer of the lesson and the AI strengthens professional judgement instead of replacing it.

In addition, we take AI literacy seriously. Every implementation of EduGPT is accompanied by training in which teachers learn to critically assess AI output and use it didactically. This is not an hour of button training, but a learning trajectory for the whole team.

And because we work for European schools, EduGPT operates fully EU-sovereign and privacy-friendly. Especially in an age of AI, that is not a detail, but a requirement.

The lesson from this research is therefore not that AI does not belong in the classroom. The lesson is that AI should never be available separately from didactics and from the professionals who work with it.

Sources

Sungu, A., Lira, B. and Duckworth, A. (2026). Generative AI Can Harm Teaching. SSRN preprint.

De Morgen (17 August 2026). Pupils find lessons less enjoyable when teachers use AI in preparation.

Image generated with AI commissioned by CivAI