An AI ban left Vrije Universiteit Amsterdam law teams last in both years of a course trial
Thibault Schrepel, who teaches law at Vrije Universiteit Amsterdam in the Netherlands, turned his Law of AI course into a three-way trial for two years: one group was kept away from AI, a second used it freely, and a third was trained to use it. The teams kept away averaged last place in 2024 and again in 2025, while the edge held by the trained group had all but vanished by the second year.
Artificial Intelligence··Evening
The course split into banned, free-use and trained groups
Thibault Schrepel teaches the Law of AI course at Vrije Universiteit Amsterdam, a university in the Netherlands, and for two years he turned it into a controlled trial. One set of students was kept away from generative tools, a second could use them with no instruction at all, and a third was taught how to prompt a model and check what came back. Teams of four or five students then had 20 minutes to rewrite one provision of the EU AI Act, the European Union's law on artificial intelligence. In the first year, 2024, the trial covered 66 students; the second year, 2025, repeated the same design with 164.[1]
Banned teams stalled while unguided teams left faulty terms in place
Groups kept away from AI mostly rewrote wording, and Schrepel reports that they ran out of ideas ten to fifteen minutes into the exercise. Groups that reached for ChatGPT, OpenAI's chatbot, without any instruction moved faster but frequently accepted its suggestions unchallenged: each of those subgroups left at least one term that was misleading or legally extraneous standing in the text it handed in. The trained group, taught to interrogate an answer before using it, delivered the strongest work in 2024.[1]
The training advantage melted away in the second year
By 2025 the three groups scored at roughly the same level, so the edge structured training had shown a year earlier no longer separated them. The teams kept away from AI still finished last on average across both years. The trial had opened with Schrepel expecting the opposite, and he says the assumption he started from did not survive it. He also sets out the study's own bounds: the sample is small, the students had chosen a course about AI, and use of the tools at home between sessions could not be checked. The result therefore speaks to two cohorts in one law faculty and does not reach across university teaching as a whole.[1]
