Establishing a sovereign infrastructure for higher education built on metacognitive student learning profiles.

UTC EduTech

Alexandre AMRANI, Université de Technologie de Compiègne Emma CHOUKROUN, Université de Technologie de Compiègne Gautier MIRALLES, Université de Technologie de Compiègne

Training / education / pedagogy

(The formatted use case description, include the three parts and the Initial Contribution, is available at: https://drive.google.com/drive/folders/11Qpv0zz_sutYagR6SXQ37PzDcRlLJ35r?usp=sharing ) Description used for application: Our project studies an AI companion integrated into a university course. Students interact with it to ask questions, generate practice exercises based on lectures and past exams, and track their progress on learning goals. The system aggregates anonymized student interactions and reports to professors which concepts cause difficulties, which questions are most frequent, and where misunderstandings emerge during the course. New description, after pivot: Most students already use AI daily. Many “tutor” platforms are being developed. The problem is not access, it is usage. Students use it to get answers, not to think, and the systems are not learning-first designed, which creates friction and inadequate model behaviors. Our project investigates a different kind of AI companion for university courses: one that knows the student, not just the course. The system builds an individual learning profile for each student, tracking their reasoning style, their recurring difficulties, and their progress over time, and uses it to generate exercises that are genuinely tailored to where they are, not just to what the syllabus says. The key design choice is this: the companion does not refuse to help, and it does not artificially limit what it can do. A system that withholds answers just pushes students back to ChatGPT. Instead, the companion engages students in the process of thinking before delivering a response e.g. asking them to articulate what they already understand, where they are stuck, and why. The answer comes, but it comes after the student has done something with their mind. On the teaching side, anonymized interaction data surfaces which concepts are causing the most difficulty, which misconceptions keep recurring, and how understanding evolves across the course. This gives professors actionable insight during teaching period and before the exam deadline. The infrastructure runs on locally-hosted, open-source models through ILaaS, keeping student data within universities control and fully GDPR-compliant. Students own their learning profile and can carry it across courses.

There is already a wide range of AI-powered learning platforms on the market. Most of them follow one of two patterns: either teachers rebuild their courses inside a proprietary AI environment to make the lectures interactive, or students chat with a bot that retrieves answers from uploaded lecture slides. Both approaches represent genuine progress over what came before, but we think they miss the deeper problem. The real issue is not that students lack access to AI. Quite the opposite: the vast majority of students already use general-purpose chatbots like ChatGPT on a daily basis [1]. These tools are fast, widely available, and more powerful than most purpose-built "educational" alternatives. A platform that just offers a slightly more structured version of the same outcome will not change behavior: students will continue turning to global black-box models, because those models simply give better answers. And "better answers" is exactly the problem. When a student receives a ready-made response, the effortful part of learning (the confusion, the false starts, the gradual construction of understanding) gets bypassed. Over time, this risks producing students who are fluent in consuming AI output but increasingly unable to reason independently [2, 3]. We believe current educational AI platforms, however well-intentioned, actively contribute to this cognitive atrophy. There is a second, related problem on the teacher's side. Professors currently have no reliable way to distinguish between a student's own thinking and AI-generated content embedded in their submissions. Plagiarism detectors are unreliable and easy to game. What teachers actually need is not detection: they need visibility into the student's reasoning process [4]. Our use case addresses both of these problems at once. Rather than building another RAG content delivery platform, we want to investigate the architecture of a thinking-first infrastructure: a system where the AI's mission is not to answer questions, but is to help students work through them and understand topics while keeping track of their progress in an individual learning profile. An innovative system that surfaces what a student actually understands and not merely what they submitted. [1] Freeman, J. Student Generative AI Survey 2025, Higher Education Policy Institute, Policy Note 61, February 2025 [2] Burley, J. AI Agents in Higher Education, Institute for Ethics and Emerging Technologies, ISBN 979-8-9879599-4-7 [3] Stadler, M. Bannert, M. Sailer, M., Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiry, Computers in Human Behavior, Volume 160, 2024, ISSN 0747-5632, https://doi.org/10.1016/j.chb.2024.108386. [4] Alshanetsky, E. When AI Dissolves Trust: Education Can Pioneer New Infrastructure. Soc 62, 982–987 (2025). https://doi.org/10.1007/s12115-025-01149-x

We are motivated by a contradiction that we experience firsthand as students: AI is supposed to support learning, but the way it is currently used in education often does the opposite. It makes getting to an answer easier, but it makes thinking harder. We want to understand how to fix this issue, and the AI Grand Challenge is the exact context to explore it seriously. Our concrete learning objectives for this project are: i. Understanding the theory behind learning profiles. We want to go beyond the idea that "everyone learns differently" and investigate what that actually means in practice: how learning science distinguishes between reasoning styles, what "scaffolding" looks like when it is done well, and how those principles can be translated into AI system design. ii. Designing a responsible infrastructure. We are specifically interested in building something that could realistically run on local, frugal models through ILaaS, rather than depending on large external APIs. This is both an ethical and a technical constraint that we want to treat as a design feature, not a limitation. iii. Operationalizing GDPR as a feature. We want to explore what it means for a student to genuinely own their data, not simply in legal terms, but in practical ones. Can a student edit the system's understanding of them? Can they carry their learning history from one course to the next? These questions sit at the intersection of governance and user experience, and we want to think through them carefully. We are also strongly motivated by the nature of the challenge itself. As computer science students, we particularly value the opportunity to brainstorm and design a project under complex and sometimes conflicting constraints, technical, ethical, and practical. This type of problem is very close to what we expect to encounter in our future careers as engineers. Beyond the intellectual challenge, this project allows us to experiment with state-of-the-art approaches in a concrete setting, and to confront them with real feedback. It is therefore not only a way to build something meaningful, but also a valuable opportunity to reflect on and improve our engineering practices in a new context.

(The formatted use case description, include the three parts and the Initial Contribution, is available at: https://drive.google.com/drive/folders/11Qpv0zz_sutYagR6SXQ37PzDcRlLJ35r?usp=sharing ) 1. What is the situation or context you are addressing? AI has become a daily tool for the vast majority of university students. Most use general-purpose chatbots like ChatGPT to get answers to course-related questions, generate content for assignments, or prepare for exams. At the same time, a growing market of purpose-built "educational AI" platforms has emerged, typically offering a chatbot interface over course materials or teacher-created content. This creates a paradox: AI is everywhere in education, yet its actual effect on learning is rarely questioned. Students have access to more assistance than ever, but the nature of that assistance (instant, frictionless answers) bypasses the cognitive effort that learning requires. Meanwhile, professors have lost visibility into how their students actually think, and existing tools give them no way to recover it. This situation is particularly critical in large university settings, where professors cannot individually track students’ understanding, making this loss of visibility even more significant. 2. What is your critical analysis of this situation? The dominant model for AI in education is content-centric: the AI retrieves or generates information, and the student consumes it. This is efficient, but it is pedagogically backwards. Research consistently shows that the effortful part of learning (confusion, trial and error, active reconstruction of understanding) is precisely what produces durable knowledge [5-9]. A system optimized to remove that friction is, in effect, optimized to impede learning. This also creates a behavioral dependency: students progressively rely on AI to bypass difficulty rather than engage with it, which fundamentally alters how they approach learning tasks. A tempting response would be to restrict what the AI can do: block answers, force students to ask differently, limit capabilities. We considered this approach and rejected it. A system that withholds help simply pushes students back to mainstream models, which are more capable and always available. Restricting the AI solves nothing; it just removes the student from a context where their learning can be tracked and supported. 3. What perspectives were discussed and how were they debated within your team? The central tension we debated was between guidance and autonomy: how much should the system steer the student, and at what point does steering become replacing? The position we converged on is that the companion should engage the student in thinking before delivering a response, not by withholding, but by asking. Where are you stuck? What have you tried? What do you already understand about this? This keeps the interaction genuinely helpful while creating a moment of reflection that a direct query to ChatGPT never would. The answer still comes; the difference is what happens before it. We also debated the scope of the learning profile: how detailed should it be, and who controls it? We agreed that a profile the student cannot see, edit, or export is a liability both ethically and in terms of trust. GDPR compliance is therefore not a constraint we work around but a design principle we want to build from. Another perspective we considered was whether forcing interaction constraints (e.g. delaying answers or enforcing step-by-step reasoning) could improve learning outcomes, but we concluded that excessive constraints would likely reduce adoption. 4. What contribution are you proposing, and under what conditions could it be implemented? We propose a thinking-first AI infrastructure for higher education, built around three interconnected layers. This proposal outlines a conceptual framework that we intend to explore and refine during the project, rather than a fully specified or already implemented system. The first is a learning profile layer. The system builds and maintains an individual profile for each student, capturing not just what they know but how they learn: their reasoning style, recurring difficulties, preferred exercise formats, and progress over time. This profile shapes every interaction (how the AI frames its questions, what exercises it generates, and how it adapts its responses) without replacing the student's effort. The second is an exercise generation layer. Rather than simple content retrieval, exercises are generated by combining course material with the student's profile. The system selects format, difficulty, and focus based on where the student actually is, not where the syllabus assumes they should be. Students can also specify what kind of practice they want, and the system will adapt accordingly. The third is a sovereign infrastructure layer. The entire system has to run on locally-hosted, open-source models through ILaaS, keeping universities in control of their data. Students own their cognitive profile: they can view it, edit it, and carry it from one course to the next. Teachers receive anonymized analytics (which concepts generate the most difficulty ,...) without access to individual student data. For example, a student preparing for an exam in mathematics could interact with the system by attempting exercises. Instead of immediately providing the solution, the AI would first prompt the student to explain their reasoning, identify where they are stuck, and then adapt its guidance accordingly. Over time, the system would detect recurring difficulties and adjust the exercises to target those specific gaps. For implementation, a proof-of-concept is feasible within our university environment using existing ILaaS infrastructure and open-source model tooling. The core technical challenge and main source of novelty lies in the learning profile: how it is defined, updated, and used to guide the system’s interactions with students. Our objective is to enable a coherent and strategic form of interaction between the AI and the student, grounded in scientifically established teaching patterns. This is the open question we intend to investigate most seriously, as it is what separates this system from a well-configured chatbot. We also aim to evaluate the effectiveness of this approach by measuring student engagement, progression over time, and their ability to solve problems independently after interacting with the system. [1] Freeman, J. Student Generative AI Survey 2025, Higher Education Policy Institute, Policy Note 61, February 2025 [2] Burley, J. AI Agents in Higher Education, Institute for Ethics and Emerging Technologies, ISBN 979-8-9879599-4-7 [3] Stadler, M. Bannert, M. Sailer, M., Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiry, Computers in Human Behavior, Volume 160, 2024, ISSN 0747-5632, https://doi.org/10.1016/j.chb.2024.108386. [4] Alshanetsky, E. When AI Dissolves Trust: Education Can Pioneer New Infrastructure. Soc 62, 982–987 (2025). https://doi.org/10.1007/s12115-025-01149-x [5] Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x [6] Karpicke, J. D., & Roediger, H. L. (2008). The critical importance of retrieval for learning. Science, 319(5865), 966–968. https://doi.org/10.1126/science.1152408 [7] Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students' learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266 [8] Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher et al. (Eds.), Psychology and the Real World (pp. 56–64). Worth Publishers. [9] Stadler, M., Bannert, M., & Sailer, M. (2024). Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiry. Computers in Human Behavior, 160, 108386. https://doi.org/10.1016/j.chb.2024.108386