A new University of Chicago survey found students misjudge classroom AI use by 30 percentage points. The misperception is doing more work than the actual use.
You can ask students what they’re doing and most of them will tell you. Ask them what their classmates are doing and most of them will be wrong.
A new Hechinger Report piece by Kristin Fasiang and Jill Barshay covers a University of Chicago survey that ran the numbers. Alex Kale’s team surveyed 338 undergraduates anonymously, with a follow-up of 98 online. Sixty percent reported using AI tools for their schoolwork. Ninety percent believed the average student on campus did.
Same school, same year. Thirty-point gap.
The researchers weren’t sure which side of the gap was wrong. Students might be underreporting their own AI use to look better, overestimating peers’ use because the loudest cases are most visible, or both. What they were sure of was that the gap exists, and that it shapes behavior. Kale put it this way: “Students don’t want to be perceived by their peers as not able to do the work. They don’t want to be perceived by their peers as dishonest. And it feels deeply personal.”
This is undergrads at an elite university. The mechanism is older.
The thirty-year-old research that explains the gap
In the 1980s, college campuses ran a study on student drinking that produced a finding now standard in social-norms research: students overestimate how much their peers drink, and underestimate how much those peers object to the drinking they see. The technical term is pluralistic ignorance. The colloquial version is “everyone but me.” Students reach for the bottle because they think it’s expected. The norm they’re trying to match isn’t real.
The intervention that eventually worked was making the actual norms visible. Tell students that most peers drink less than they assume. Show the numbers. The misperception softens, and behavior shifts.
The UChicago finding maps onto that frame. AI in college runs on the same social mechanism: behavior driven by what students assume others are doing. Use is high. Perceived use is higher. The gap is its own engine.
Why this matters for high school more than college
High school students are more peer-sensitive than undergrads, not less. The signals about what other kids are doing arrive louder, the consequences of being out of step feel sharper, and the pathways for verifying what’s happening are narrower. The teacher is one of the few people in their lives who sees the full distribution of behaviors. The kids see only what their peers tell them, and those self-reports get filtered through the same social pressure that produced the gap in the first place.
If 60% of elite-college undergrads use AI and 90% think their peers do, the equivalent numbers in a tenth-grade classroom likely move in the same direction with more amplification. Most students using AI less than they let on. Most students thinking everyone else uses it more than they do.
That has a practical consequence. Every kid who hesitates to outsource a draft is doing so under the impression that they’re alone in that hesitation. Most of them aren’t.
I see this in my own classes. The students who tell me they “couldn’t have done it without ChatGPT” are sometimes the strongest writers in the room: kids who could have produced the same work without help, but assumed everyone else was using the tool and felt foolish not to. The signal they were reading was the noisiest one. The noisiest signal isn’t the truest one.
What this means for teachers
The most useful intervention is the cheapest one. Surface the actual norms.
You don’t need a controlled study. A one-question anonymous survey at the end of a unit, asked clearly, gets you most of the way. “On the last writing assignment, how much of your draft came from AI? Be honest. Results will be shared anonymously.” Then share them. The kids who think they’re alone in writing their own drafts see that they aren’t. The kids who routinely outsource see that the social cover they assumed is thinner than they thought.
This is information correction, not surveillance. The same intervention that softened college drinking by showing students the real numbers can soften the AI-everyone-is-doing-it pressure by showing students who is and isn’t.
The harder thing is to do this without turning it into a shame ritual. How you set it up matters. You aren’t catching anyone. You’re showing the class what the class looks like, so the kids who want to do their own work can stop thinking they’re swimming upstream.
There’s an honest caveat worth flagging. The UChicago survey is one elite-college sample at one moment, and the researchers themselves can’t yet resolve whether students are lying down (underreporting their own use) or imagining up (overestimating peers’). Both interpretations lead to the same intervention, but the underlying social dynamics differ in ways that probably matter for what you do next. Treat the 60/90 split as directional, not as the floor of K-12 reality.
Takeaway for Teachers
This week, run a one-question anonymous survey on your last writing assignment. “Roughly what share of your draft came from AI?” Five buckets: none, a little, some, a lot, basically all of it. Share results back to the class without commentary. Let them see the actual distribution.
If you teach the same students next year, run it again. The first run gives you a baseline. The second tells you whether sharing the truth changed anything.
The number itself does some of the work.
If you’re thinking through questions like this one — what your students assume about each other, and how those assumptions shape what they do in your classroom — my book goes deeper. The AI Doesn’t Know Your Students is available on Amazon and at shouldiuse.ai/book.
David Jacobson is a high school history teacher in Shanghai. He writes about AI and education at shouldiuse.ai and is the author of The AI Doesn’t Know Your Students.
