Two-year university study finds banning AI from classrooms leaves students worse off

Details

Law professor Thibault Schrepel ran a randomised experiment across 2024 (66 students) and 2025 (164 students) in his Law of AI course at Vrije Universiteit Amsterdam. Three groups were given 20 minutes to improve an EU AI Act provision, graded on substance, clarity, proportionality, and innovation: one group had no AI access, one received embedded ChatGPT suggestions with no guidance, and a third received hands-on training in legal prompt engineering and output verification. The no-AI group consistently ran into what Schrepel labels 'idea exhaustion' after 10-15 minutes, defaulting to minor wording tweaks. The unguided AI group accepted ChatGPT suggestions largely at face value — every subgroup retained at least one legally flawed or extraneous term from the AI output. Only the trained group engaged in genuine back-and-forth, testing phrasings and probing the substantive legal questions. In 2024 the trained group outperformed the others significantly, particularly on a take-home exam requiring deeper legal reasoning. By 2025 that advantage had nearly disappeared as students arrived with stronger everyday AI fluency. One result held firm across both years: the no-AI group finished last. Schrepel concludes that AI bans produce worse average outcomes than even unguided use, and that structured training delivers the largest benefit for students with lower prior AI experience. He notes ethical use and legal responsibility still require explicit instruction regardless of AI access.