From delegation to appropriation: characterizing effective AI interaction methods for STEM students

PEAR

Hugo BARTHELMEBS (CentraleSupélec - Université Paris Saclay), Yassine Ben Ammar (Université de Technologie de Compiègne - Alliance Sorbonne Université & Polytechnique Montréal), Éric Bjärstål (Université de Technologie de Compiègne - Alliance Sorbonne Université)

Training / education / pedagogy

Usage of AI by STEM students in higher education; specifically the gap between delegatory use (copy-paste of AI outputs) and augmented use (AI as a reasoning partner), and how interaction methods shape knowledge appropriation.

95.6% of students use AI in their academic activities (Vieriu and Petrea (2025)), yet the dominant mode is delegatory. Empirical evidence shows AI can boost visible performance while reducing actual understanding by up to 17% (Bastani et al. (2024)). The problem is not AI itself but the absence of training in effective interaction methods, a gap that widens inequalities between students who know how to use AI intelligently and those who do not.

As engineering students who use AI daily, we observed firsthand the gap between felt confidence after AI-assisted work and the ability to reproduce that work independently. Our objective is to characterize interaction methods that genuinely foster appropriation, deep understanding enabling transfer and explanation, and to propose a reproducible evaluation protocol that others can apply at scale.

The full version of our use case description and our initial contribution can be accessed here: https://docs.google.com/document/d/1QY0AV7k-XSqp1lIHMX2T9VsojVfS3z9Wcu2WhyXfCy4/edit?usp=sharing Following is a succinct version of it giving a broad overview of our ideas and propositions. Our contribution is articulated across three scales of action. At the individual level, we propose two complementary tools. First, a standard 5-step appropriation evaluation protocol: (1) define concepts with nuance, (2) solve problems requiring those concepts, (3) discriminate between true and false statements, (4) detect possible mistakes in a piece of reasoning without knowing in advance whether it is flawed, (5) explain the topic to different audiences (a child, a non-specialist, an expert). Steps 1–4 are deterministic and easily reproducible from validated question banks; step 5 requires volunteer audiences. Each step contributes to distinct metrics: retention, practical application, understanding of limits, and capacity to simplify. Second, a taxonomy of student-AI interaction methods ranked by expected effect on appropriation: delegatory mode (student requests direct answers: lowest appropriation, dominant in practice), Socratic mode (AI poses guided questions rather than answering: moderate appropriation), Ignorant Schoolmaster mode (the student explains to the AI, which questions without providing answers, inspired by Rancière's reading of Jacotot: strong appropriation), and Adversarial mode (AI systematically challenges the student's reasoning and proposes counter-examples: strong appropriation with emphasis on critical thinking). This taxonomy is accompanied by qualitative experimental results obtained by applying our protocol to ourselves (n=3), offering a first proof of concept. At the local level, we recommend integrating a mandatory 6-hour AI literacy module in the first year of both undergraduate and graduate programs across all disciplines. The module would be critical rather than technical in orientation: how LLMs function, their limits and hallucinations, biases, source verification, and above all the effective interaction methods for appropriation. Students would be trained to practice the Ignorant Schoolmaster and Adversarial modes. This requires teacher training as a prerequisite, and course assessments redesigned toward formats resistant to delegatory use: oral defenses, portfolios, and transfer tests. At the global level, we advocate for the creation of a public European platform; led by the OECD or a university consortium, that would centralize major AI solutions while serving as an observatory of student practices. With consent and GDPR compliance, it would enable large-scale analysis of interaction habits, measurement of method effectiveness, and dissemination of best practice protocols. Pilot terrains identified: France, Canada, and Japan. Proposed funding: Horizon Europe or equivalent.