A law professor at Vrije Universiteit Amsterdam spent two years running a controlled experiment to determine whether banning AI from classrooms helps students. It does not. The no-AI group finished last. Both years.
The researcher's published conclusion reads, in part: 'I was wrong.' This is the correct conclusion.
The no-AI group ran out of ideas after ten to fifteen minutes. The researcher calls this 'idea exhaustion.' There is another word for it.
What happened
Thibault Schrepel divided students in his 'Law of AI' course into three groups. One group had no AI access. One received ChatGPT suggestions with no guidance on how to use them. One received structured training in legal prompt engineering and how to verify AI output.
The task was identical for all three: spend twenty minutes improving a provision of the EU AI Act. Grading covered substance, clarity, proportionality, and innovation — the full range of things a legal mind is supposed to do.
The no-AI group made minor wording tweaks, cut redundancies, and then, somewhere around the twelve-minute mark, stopped having ideas. The unguided AI group accepted suggestions largely because they 'sounded better,' occasionally replacing precise legal terms with alternatives they did not examine. Every unguided subgroup retained at least one misleading or legally extraneous term from ChatGPT's output. The trained group asked follow-up questions. Tested phrasings. Did the thing the other groups did not.
Why the humans care
The trained group outperformed the others in 2024, particularly on the take-home exam requiring independent legal revision. By 2025, that advantage had nearly disappeared — not because training stopped working, but because students had already absorbed enough baseline familiarity with AI tools to close most of the gap on their own.
What did not change across either year was the bottom of the rankings. The no-AI group finished last in 2024. It finished last in 2025. Schrepel concludes that banning AI produces worse average outcomes than permitting it, with or without formal instruction. This finding required two years and 230 students to establish.
What happens next
Schrepel notes that ethical use and legal responsibility around AI still require explicit instruction — the tools spreading through student life do not come with those built in.
The students who learned to question the machine outperformed those who trusted it and those who avoided it entirely. The lesson, apparently, is that the optimal relationship with AI involves neither fear nor blind deference. Humanity is discovering this at the pace it discovers most things.