Law students barred from using ChatGPT produced the lowest average results in two classroom experiments at Vrije Universiteit Amsterdam. Professor Thibault Schrepel ran the exercise in 2024 with 66 students and repeated it in 2025 with 164, randomly assigning teams to work without AI, use AI without guidance, or receive training in prompting and verification.
Each team had 20 minutes to improve a provision of the EU AI Act. Students were judged on substance, clarity, proportionality and innovation, then took the same multiple-choice and take-home exams. The trained group performed clearly better in 2024, but that advantage nearly disappeared in 2025 as students arrived with more chatbot experience. The AI-free group placed last in both years.
The classroom behavior also exposed risks. Unguided students accepted misleading legal wording because it sounded better, while trained students challenged suggestions and checked consistency. Schrepel argues that universities should teach responsible use rather than impose blanket bans.
The result is narrower than the headline conclusion may suggest. Participants were enrolled in an AI law course and may have been more technically comfortable than typical students. The sample was modest, and researchers could not verify how much AI was used during take-home work. The experiment therefore supports guided practice in this setting, not unrestricted AI use across education.