How Students Use AI: Evidence from an AI Tutor at University

Abstract

We study the extent, nature, and malleability of students' AI use at a German distance-learning university. We link more than 100,000 conversations with a GPT-based tutor chatbot to administrative records for more than 7,000 students, and run a preregistered randomized encouragement campaign. Baseline adoption is selective, concentrated among students with prior AI experience and higher digital competence. Supervised text classification shows that most requests concern concept explanations or task solving, while higher-order uses similar to cognitive activation are rare — a pattern closer to cognitive substitution than to complementarity in skill formation. The degree of cognitive delegation varies widely across students but is highly persistent within them. The campaign raises adoption among newly enrolled students, with no detectable effect among incumbents. Preliminary findings on downstream outcomes indicate higher study progress among new enrollees and lower dropout among incumbents, with no effects on grades. Joint paper with Philipp Höllermann, Ingo Isphording & Sebastian Riedmiller.