The rapid adoption of Large Language Models (LLMs) in education offers new opportuni-
ties to support student learning, while also raising critical concerns about how AI tutors can
be aligned to foster durable learning gains (Kasneci et al. 2023). Unlike professional users of
conversational agents, learners cannot be expected to monitor or correct whether an AI consis-
tently maintains a high-quality pedagogical role. General-purpose systems such as ChatGPT,
which rely on user initiative and precise prompting, are prone to factual inaccuracies and role
drift, making them only partially suitable as tutors (Bastani et al. 2025). An Intelligent Tu-
toring System (ITS) must therefore sustain its pedagogical framing autonomously through a
reliability-oriented architecture that secures the educational context independently of learners’
input discipline.
This raises the research question: How should such an ITS be designed to support stu-
dents’ learning? To contribute to this question, this paper introduces AIDu, a concept for
an exploratory, self-directed AI tutoring system for upper secondary education, grounded in
epistemic AI-tutoring. In contrast to broadly labeled Socratic tutors that scaffold learning with
hint-like questions toward predefined answers - potentially limiting deeper learning - AIDu aims
to foster epistemic engagement. It provides open-ended, non-directive feedback that promotes
exploration, hypothesis generation, and reflection, rather than steering learners along a fixed
solution path.
We outline AIDu’s design principles, present a prototype, and report findings from two
trials. As LLMs perform most reliably when context is concretely defined, we embedded AI
dialogues in authentic, virtual learning environments such as interactive applets. This strategy
addresses role drift and hallucinations while combining subject-specific structures with active,
self-directed engagement and cognitively activating processes (Kaldaras et al. 2024).
An exploratory study with 8 students and a subsequent larger study with 117 students
provided usability feedback via structured polls, complemented in the second iteration by
qualitative discourse analysis of 38 student-AI dialogues. Results indicate that epistemic AI-
tutoring, when embedded in authentic environments, helps stabilize tutor roles and supports
reflective, self-directed learning in secondary schools.
References
Bastani, Hamsa et al. (2025), “Generative AI without guardrails can harm learning: Evidence
from high school mathematics,” in: Proceedings of the National Academy of Sciences 122.26,
e2422633122.
Kaldaras, Leonora et al. (2024), “Employing technology-enhanced feedback and scaffolding to sup-
port the development of deep science understanding using computer simulations,” in: International
Journal of STEM Education 11.1, 30, issn: 2196-7822.
Kasneci, Enkelejda et al. (2023), “ChatGPT for good? On opportunities and challenges of large
language models for education,” in: Learning and Individual Differences 103, 102274.
Puntambekar, Sadhana and Roland Hubscher (2005), “Tools for scaffolding students in a complex
learning environment: