Thibault Schrepel, a law professor at Vrije Universiteit Amsterdam, entered his two-year experiment convinced that unguided AI use would harm legal education. He was wrong. After splitting students in his “Law of AI” course into three distinct groups, the data revealed a counterintuitive truth: students forbidden from using ChatGPT consistently produced the weakest work. This isn’t just a theoretical debate; it’s a practical failure of prohibition. The no-AI group suffered from what Schrepel terms “idea exhaustion,” running out of substantive improvements after just ten minutes and resorting to minor wording tweaks. Meanwhile, their peers with access to AI kept generating options, even if they didn’t always understand them.
The study, which expanded from 66 participants in 2024 to 164 in 2025, tested three approaches: a total ban, unguided access to ChatGPT suggestions, and structured training in prompt engineering. The trained group initially dominated, scoring well above the others on complex take-home exams. But by year two, that advantage nearly vanished. Schrepel attributes this leveling effect to growing everyday familiarity with chatbots. Students are learning to spot AI weaknesses through trial and error, even without formal instruction. The unguided group actually outperformed the banned group, suggesting that hands-on experience teaches critical evaluation faster than abstinence does.
This creates a sharp conflict with institutions like UC Berkeley Law, which has banned AI from most graded work to protect core thinking skills. Schrepel’s findings suggest that approach may be backward. By removing the tool, universities aren’t preserving rigor; they’re limiting output. He argues that traditional master’s theses, often reliant on literature reviews that AI can now generate in minutes, need rethinking. Programs should shift toward practice-oriented or empirical work where AI use is part of the grading criteria rather than a cheating violation.
Faculty development remains the biggest bottleneck. Many instructors feel ill-equipped to guide these tools, leading to blanket bans as a default safety measure. But the evidence shows that skipping AI instruction is a missed opportunity. While the study has limits—small sample sizes and self-selected tech-savvy law students—the trend is clear. Prohibition doesn’t build better thinkers; it just builds slower ones. Universities must decide if they want to train students for the world as it exists, or as they wish it were.
