Two-year university study finds banning AI from classrooms leaves students worse off
decoder ·
A two-year study by law professor Thibault Schrepel at Vrije Universiteit Amsterdam examined how an AI ban, unguided AI suggestions, and structured training affected students in his Law of AI course. Students were randomly assigned to three groups: one barred from ChatGPT, one given ChatGPT-generated revision suggestions without guidance, and one trained in legal prompt engineering and checking AI output. Teams of four or five had 20 minutes to improve an EU AI Act provision, with grading for substance, clarity, proportionality, and innovation, and all took the same multiple-choice and take-home exams. The study ran in 2024 with 66 students and in 2025 with 164. The no-AI group finished last both years, often running out of ideas after 10-15 minutes, while the unguided AI group accepted suggestions uncritically yet slightly outperformed the no-AI group on exams. The trained group scored well above the others in 2024, especially on the take-home exam, but the advantage nearly disappeared in 2025, which Schrepel attributes to growing chatbot familiarity. Schrepel writes he was wrong to assume AI only helps with structured training, concludes that banning AI leads to worse average outcomes, and recommends that universities avoid blanket bans, invest in faculty AI skills, and rethink the master's thesis and pure literature reviews; he notes that UC Berkeley Law's ban on AI from nearly all graded work conflicts with his findings. He acknowledges limitations including small samples, tech-savvy students in an AI course, and no way to verify take-home AI use, while other research cited found AI raised grades on unsupervised work but proctored exams showed no comparable shift and a China K-12 study saw exam performance drop up to 24 percent where AI replaced independent thinking.
decoder