Two-Year Classroom Study Finds AI Bans Hurt Student Performance, Researcher Reverses Stance

A law professor who spent two years comparing three classroom approaches to artificial intelligence has concluded that banning the technology outright leaves students worse off than letting them use it. According to a report from The Decoder, the study tracked an AI ban, unguided AI use, and structured training across two academic years, and the group without any access to AI tools finished last in both. The finding lands as universities worldwide continue to rewrite course policies around generative models, often with little experimental evidence to guide them.

The design put three cohorts side by side: one barred from AI, one given access with no instruction on how to use it, and one that received structured training on working with the tools. The researcher, a law professor, had expected the unguided group to fare worst, assuming students would lean on AI in ways that undercut their own learning. That assumption did not survive the data. The no-AI cohort placed at the bottom in each of the two years measured, while the structured-training group posted the strongest results.

The reversal is notable because it comes from the researcher directly. In the write-up, the professor states plainly, "I was wrong," a concession that cuts against the instinct driving many current bans. The admission matters for how the result will be read: this is not an AI advocate arguing from the outside, but a faculty member reporting that an experiment contradicted his own prior position.

The result also complicates the framing that has dominated campus debates, in which the real risk is seen as students using AI without guardrails. If unguided access still outperformed no access at all, then the harm of prohibition may exceed the harm of unstructured use, at least on the outcomes this study tracked. The study does not claim that structured training is optional or that any AI use is equally effective, but it does undercut the case for blanket bans as a default policy.

What remains unresolved is how far the finding travels. The experiment was run in a law course over two years, so its reach to other disciplines, class sizes, and assessment formats is untested. The professor's own shift suggests the more useful question is no longer whether to permit AI in classrooms, but what kind of instruction around it produces the best results. Institutions weighing bans now have a two-year data point arguing the other way.