More than 80% of Middlebury University students use generated AI for their coursework, according to a recent survey conducted with colleagues and economist Zara contractors. This was one of the fastest technology adoption rates on record, far surpassing the adoption rate of 40% of US adults, and less than two years after ChatGPT's release.
We investigated only one university, but our results are consistent with similar research and provide new photos of the use of technology in higher education.
Between December 2024 and February 2025, we surveyed over 20% of Middlebury College's student organization (634 students) and published the results in a working paper that has not yet experienced peer reviews.
What we find challenges the panic-driven narrative of AI in higher education, suggesting that institutional policies should instead focus on how AI is used.
Contrary to the surprising headlines suggesting that “ChatGpt unravels the entire academic project” and “AI fraud is getting worse,” students have found that primarily use AI to enhance learning rather than avoiding work.
When I asked students about 10 different academic uses of AI, I was top of the list, explaining the concepts, from concept explanations and proofreading summary, writing programming code, and of course writing essays. Students frequently described AI as “on-demand tutors.” This is a resource that is especially valuable when opening hours are unavailable or when you need help late at night.
We grouped AI usage into two types, describing “enhancement”, describing use to enhance learning, and describing “automation” of use to generate work with minimal effort. 61% of students using AI use these tools for augmentation purposes, while 42% use them for automated tasks such as writing essays and generating code.
They also showed judgment, even if students used AI to automate tasks. Open-ended responses said that when students automate their work, they often have low-stakes tasks, such as bibliographic formatting or drafting regular emails, rather than the default approach to completing meaningful coursework during crunch times like exam week.
Of course, Middlebury is a small liberal arts college, with a relatively large majority of wealthy students. How about anywhere else? To investigate, we analyzed data from other researchers covering over 130 universities in over 50 countries. The results reflect Middlebury's findings. Globally, students using AI tend to be more likely to use it to increase coursework rather than automate it.
But do you need to trust that students will teach you how to use AI? An obvious concern with research data is that students overreport legitimate uses that get explanations while underestimating what they deem as inappropriate, like writing an essay. To examine the findings, they were compared with data from humanity from AI companies. This analyzed actual usage patterns from the chatbot's Claude AI university email address.
Anthropic's data shows that “technical explanations” represent key uses, consistent with the findings that students use AI to explain concepts most often. Similarly, Anthropic found that by designing practice questions, editing essays, and summarizing the material, it constitutes a significant portion of student use that matches the results.
In other words, our self-reported survey data matches the actual AI conversation log.
Why is it important?
As author and academic Hua HSU recently stated, “There are no reliable numbers for American students to use AI. It's just a story about how everyone is doing it.” These stories tend to highlight extreme examples, like Columbian students who used AI to “deceive almost every task.”
However, these anecdotes can confuse widespread adoption with universal misconduct. Our data confirms that the use of AI is indeed widespread, but students do not primarily use it to enhance learning and replace it. This distinction is important. By portraying all AI use as fraud, vigilance coverage can normalize academic injustice, and responsible students feel naive in obeying when they believe that “everyone else is doing it.”
Furthermore, this distorted image provides biased information for university administrators who need accurate data on actual student AI usage patterns to create effective evidence-based policies.
What's next?
Our findings suggest that extreme policies, such as the ban on blankets and unlimited use, are risky. The ban disproportionately hurts students who benefit most from AI's individualized teaching capabilities, creating unfair benefits for rule breakers. However, unlimited use may allow harmful automation practices that can undermine learning.
Instead of a policy of all sizes, our findings will lead us to believe that institutions should focus on helping them distinguish between potentially harmful and beneficial AI use. Unfortunately, research on the real impact of AI on learning remains in its early stages. No studies have systematically tested how different types of AI use affect students' learning outcomes or whether the impact of AI is positive for some students but negative for others.
Until that evidence becomes available, everyone interested in how this technology is changing must use their best judgment to determine how AI can promote learning.
The research summary is a short view of interesting academic work.
