When is generative AI truly educational? Seven characteristics from research and practice
Recently, Pascal Mariany received a question in an education community that has stayed with him. What must generative AI fulfil in order to be truly educational? And not just another chatbot with an education label, but AI that genuinely advances learning.
A justified question. Almost every provider now calls their tool educational. A study mode here, a student version there. But what makes the difference between AI that teaches and AI that simply does the work?
In this piece, Pascal outlines what research and practice tell us about this. Especially in an era of AI, that question deserves a substantiated answer.
Educational value lies in the design, not in the label
The most important insight first. The same language model can either harm or enhance learning. That depends on how it is set up and how you use it.
The most convincing evidence comes from Hamsa Bastani and colleagues (PNAS, 2025). They gave nearly a thousand secondary school students in Turkey access to GPT-4 for mathematics. One group received a standard chat interface, the other a tutor version that gave hints instead of answers.
During practice, the group with the ordinary chatbot scored 48 per cent better than students without AI. But when the AI was removed, they performed 17 per cent worse on the test than students who had never used AI. They had used the chatbot as a crutch. In the tutor version, this damage was largely avoided.
That is why Pascal looks at educational AI through the lens of HEAR, the framework for human-centred work with AI from my book The Art of Living Together with AI. HEAR has four layers. Human Intent concerns the purpose, which always lies with the human. Expressive Interaction is about the language between human and system. AI Mediation is about what happens between question and answer. Reasoned Responsibility is about who bears the judgement. You will see these layers reflected below.
Seven characteristics of educational AI
1. The thinking remains with the student
Educational AI does not provide ready-made work. It first asks what the student has already tried, where they are stuck, and what the next step could be. Only then does it provide a hint, a framework, or an example.
This is precisely the layer of AI Mediation. What happens between question and answer determines whether learning takes place. Michael Gerlich (2025) found a clear link among 666 participants between frequent AI use and reduced critical thinking, via the outsourcing of cognitive work. Good educational AI consciously slows down that outsourcing.
2. The design is based on learning science
Good educational AI applies principles we have known for decades. Think of active learning, dosing cognitive load, prompting recall from memory, and providing feedback on the process rather than just the end product.
Hattie and Timperley (2007) already described that feedback only works if it answers three questions. Where am I going, where am I now, and what is my next step? An AI that only returns a grade or an improved text skips those questions.
That design pays off. Gregory Kestin and colleagues (Scientific Reports, 2025) built an AI tutor for physics at Harvard on the same didactic principles as their lessons. With it, students demonstrably learned more in less time than in a lesson with active learning, and also felt more engaged.
3. The learning objective comes first
Educational AI knows what it is being used for. It is tailored to the learning objective, the level, and the curriculum, and the teacher can set that context themselves.
This is where Human Intent comes in. The purpose comes from the human, not the system. An AI that gives the same helpful answer to every question, regardless of whether it is a test, an exercise, or an exploration, completely misses that purpose.
4. The teacher remains the director
The best results, Pascal observes, are where AI enhances the teacher rather than replaces them. Rose Wang and colleagues (Stanford, 2024) studied Tutor CoPilot with 900 tutors and 1,800 students. The AI gave the tutor suggestions during the lesson, but the tutor remained in charge.
Students achieved mastery of the material 4 percentage points more often, and among less experienced tutors even 9 percentage points more. Tutors asked more guiding questions and gave away the answer less often.
This is Reasoned Responsibility. The judgement about what a student needs remains with the professional. The AI supports that judgement but does not take it over.
5. The AI is honest about uncertainty and sources
Language models formulate convincingly, even when they are wrong. That is the pitfall of Expressive Interaction. The system sounds confident, but means nothing.
Research from Microsoft Research (Lee et al., 2025) shows that people who have a lot of trust in AI think less critically about the outcome. Educational AI must therefore calibrate that trust. It shows where its knowledge comes from, acknowledges doubt, and invites the student to check.
6. Safe and age-appropriate
Students under eighteen are a vulnerable group. Privacy, data storage, and age limits must therefore be in order before the didactics even come into play.
The AI Assessment Framework for Primary and Secondary Education from SIVON and Kennisnet (April 2026) is a good starting point for this. And if you use AI to assess or determine access to education, it falls under the high-risk rules of the European AI Act, which will apply from December 2027.
7. Measure learning, not satisfaction
How well students perform with AI says little about what they have learned. Bastani’s research painfully illustrates this. The best practice scores went hand in hand with the worst test results.
If you want to know whether an AI is educational, measure what students can do independently afterwards. That requires a pre-test and a post-test, and the courage to abandon a tool if it does not help.
Five questions for those choosing or building educational AI
Does the AI give answers or does it ask questions in return?
On which didactic principles is the design based, and can the provider demonstrate this?
Can the teacher set the learning objective, the level, and the boundaries themselves?
How does the AI deal with uncertainty, sources, and student data?
Is there research into what students can do independently after using it?
How we apply this in practice
These insights have been the reason for us to build our own critical thinking layer into EduGPT. This sits between the student’s question and the model’s answer. It does not provide ready-made work, but asks what the student has already done and helps with hints and a framework. Guiding instead of delivering.
It remains a choice for the teacher or course. Sometimes you do want a direct answer, for example when preparing a lesson. Educational does not mean that the AI always asks back, but that its use matches the learning objective.
Pascal is curious about your perspective. Which characteristics do you still miss, and what do you see happening in your own class or course? Let’s have the conversation together, because information is everywhere, but guidance is not.
Sources
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2025). Generative AI without guardrails can harm learning. Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. doi.org/10.1073/pnas.2422633122
Gerlich, M. (2025). AI tools in society. Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), 6. doi.org/10.3390/soc15010006
Hattie, J. and Timperley, H. (2007). The power of feedback. Review of Educational Research, 77(1), 81-112. doi.org/10.3102/003465430298487
Kestin, G., Miller, K., Klales, A., Milbourne, T. and Ponti, G. (2025). AI tutoring outperforms in-class active learning. An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. doi.org/10.1038/s41598-025-97652-6
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R. and Wilson, N. (2025). The impact of generative AI on critical thinking. Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, 1-22. doi.org/10.1145/3706598.3713778
Mariany, P. (2026). The Art of Living Together with AI. Van Duuren Media.
SIVON and Kennisnet (2026). AI Assessment Framework for Primary and Secondary Education, version 1.0.
Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S. and Demszky, D. (2024). Tutor CoPilot. A human-AI approach for scaling real-time expertise. arXiv, 2410.03017. doi.org/10.48550/arXiv.2410.03017
Image for illustration, created with AI