Why It Matters
As colleges and universities nationwide wrestle with how to respond to artificial intelligence in the classroom, new research out of Vermont offers data that complicates simple narratives around either banning or embracing the technology. The findings suggest that blanket policies may inadvertently disadvantage students who use AI responsibly.
What Happened
Germán Reyes, an assistant economics professor at Middlebury College in Vermont, and his colleague Zara Contractor carried out a two-part study examining how students actually use AI for academic work and what effect that use has on their learning outcomes.
The first phase was a survey conducted between December 2024 and February 2025, in which Middlebury students were asked about their AI habits in academic settings. The researchers cross-checked those survey responses with data from Anthropic, the company behind the Claude AI platform, using records tied to college email addresses. They also compared the Middlebury results to broader data drawn from more than 50 countries.
The second phase, run in spring 2025, was a controlled experiment. Participants were randomly divided into two groups — one had access only to standard online tools, while the other could also use AI. Both groups researched topics including CRISPR gene-editing technology and wrote essays. A week later, participants returned to write a second essay on the same subject and take a test, this time without any tools.
Reyes and Contractor categorized student behavior into two distinct patterns: augmentation, where students used AI to deepen their understanding of material, and automation, where students delegated the work to AI rather than engaging with it themselves.
By the Numbers
80% of Middlebury students reported using AI for academic purposes, according to the survey. Roughly 70% of participants in the AI-enabled group adopted the tool during the experiment.
The results diverged sharply based on how students used it. Automation users produced notably stronger first essays but performed significantly worse than their peers on the second essay and the no-tool test — suggesting short-term gains at the expense of actual learning. Augmentation users showed a more modest lift in the first round but demonstrated stronger retention and performance in the second round, when tools were removed.
Reyes has secured funding to repeat the survey in 2026, with additional questions incorporated to refine the research.
Key Takeaway
The distinction between augmentation and automation is central to what the researchers say institutions should consider before setting policy. Reyes argued that a blanket prohibition could penalize students who are genuinely using AI to learn more effectively. “If some students are using it to learn more and you decide to ban AI because you think it has no room in college to help students learn, then you’re inadvertently going to be harming students who benefit from the tool,” he said.
On the automation side, the tradeoff is equally clear. “When you have access to the tool and use it to do the work for you, that provides large, short-term gains,” Reyes noted. “But that comes at the cost of lower learning in the long run.”
Zoom Out
The Middlebury findings arrive as institutions across the country are still searching for a coherent framework for AI in higher education. Some have moved toward outright bans; others have encouraged integration. The study’s data — validated against global figures and cross-referenced with usage records from an AI developer — adds empirical weight to a debate that has largely been driven by instinct and anecdote.
Vermont’s technology sector has been growing in parallel, with companies like Beta Technologies attracting skilled workers from smaller regional aviation firms, illustrating a broader pattern of AI and advanced technology reshaping both industry and education in the state.
What’s Next
Reyes plans to conduct a follow-up survey in 2026 with expanded questions, building on the initial findings. The next phase is expected to shed additional light on whether patterns in AI use are shifting as the technology becomes more embedded in academic life, and whether the augmentation-versus-automation divide holds across different disciplines and student populations.