How can I use AI as a tutor to support my students' learning?

I don't know if I want to convince you (and I don't know if I am convinced myself), but I recently read several studies about AI tutors that made me feel slightly more positive towards the idea. The way that AI tutors were used in those studies was to set them up in ways that minimise hallucinations (which can never be excluded completely, but the GenAI can be trained on the exact materials we want), do not give students complete solutions but rather help them with the next step, and are generally supportive and encouraging of a growth mindset. In those studies, students did learn as much or even more from the AI tutors as they did in in-person active learning control groups, and they appreciated the self-paced nature. Kestin et al. (2025) might be an interesting article to read in that context.

The conceptual paper by Brcic & Frljic (2026) gives a nice framework for when and how to use, and importantly when not to use AI on a very small-scale level, definitely worth having a look at that article, especially their Table 1 and 2!

What all the studies I read warn against, though, is using unguarded GenAI as tutor, that seems to get used too often as a crutch and thus be counterproductive to learning.

And what I think is way under-discussed is how to use AI tutors in ways that don't discriminate students (as in "you are someone who should really work with the AI tutor so you get up to speed before you come to class", or "if you cannot afford a human tutor, use the free AI one!").

One use case that I liked, described by Wang et al. (2024), was using AI not to tutor students directly, but to support human tutors by providing them with suggestions, combining content and good pedagogical practice, for how to react to students. Of course that might also be a slippery slope of the human tutor more and more uncritically relying on the AI suggestions, which might have unintended side effects on both the human tutors and their students, but that was not explored in that study.

So in a nutshell: AI tutors might be worth experimenting with!

---

Brcic, M., & Frljic, S. (2026). The Effortless Trap: Productive Struggle, AI, and the Illusion of Learning. arXiv preprint arXiv:2606.26181.

Kestin, G., Miller, K., Klales, A. et al. (2025) AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Sci Rep 15, 17458. https://doi.org/10.1038/s41598-025-97652-6

Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., & Demszky, D. (2024). Tutor copilot: A human-ai approach for scaling real-time expertise. arXiv preprint arXiv:2410.03017.

Like Mirjam, I'm cautious about this, but I do think that some scaffolded use of GenAI might be useful, particularly for learning concepts. I avoid the term "AI tutor" because a tutor must be human, but "AI exercises" might be ok. Limiting the available training material to relevant course materials is possible with the use of agents, which you can do with most current systems (writing this in 2026 - and the use of agents will probably get easier as long as costs don't get prohibitive). With an agent, you upload the materials you want the agent to refer to, and set parameters. For instance, you could say that the agent can't give factual answers, but should only ask questions to help the student to work through the concept. This may not be fully reliable, but it can help to avoid dependent behaviours in students. (Or be very frustrating for them - always check in with them about any learning materials you have produced!). As with any kind of learning materials, the real value comes when they are discussed in class with peers and with you, so make sure there is space for this, and the products are not just used for what used to be called FOFO in 1990s distance learning development (the last three letters stand for 'Off and Find Out') - see Kumar et al, 2024.

There is a section in the open access book I co-authored with Sam Illingworth on personalised learning (part 2 in chapter 4), which also gives some background on so-called Intelligent Tutoring Systems, if you want to look at some practical suggestions.

Mirjam has already cited Kestin et al above, and there is also a preprint here: Chung, A. T.-H., Zhang, B., Kung, L.-C., Bastani, H., & Bastani, O. (2026). Effective Personalized AI Tutors via LLM-Guided Reinforcement Learning. Available at SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6423358

Kumar, H., Musabirov, I., Reza, M., Shi, J., Wang, X., Williams, J. J., Kuzminykh, A., & Liut, M. (2024). Guiding Students in Using LLMs in Supported Learning Environments: Effects on Interaction Dynamics, Learner Performance, Confidence, and Trust. Proc. ACM Hum.-Comput. Interact., 8(CSCW2), Article 499. https://doi.org/10.1145/3687038

And the full GenAI and HE book is here, free to download Illingworth, S., & Forsyth, R. (2026). GenAI in Higher Education: Redefining Teaching and Learning. Bloomsbury. https://www.bloomsbury.com/uk/genai-in-higher-education-9781350535787/