The student AI-use numbers every educator should know before setting policy

Why numbers, not anecdotes

Most classroom AI policy is written in response to an incident — one shocking essay, one viral story. The survey data tells a different and more useful story, and it fits in five numbers. Each is sourced and quoted exactly in the companion review, student AI-use surveys — what the numbers say; this page is the educator's version: the numbers, then what they change about policy.

The five numbers

1. 88% — students using generative AI for assessments. HEPI/Kortext's 2025 survey of 1,041 UK undergraduates: use for assessments jumped "from 53% last year to 88% this year," and overall AI use hit 92%. The most common uses are explaining concepts, summarising articles, and suggesting research ideas. Whatever your policy says, this is the baseline it regulates: AI use for assessed work is now normal behavior, not deviant behavior.

2. 26% — US teens using ChatGPT for schoolwork, double the prior year. Pew's survey of 1,391 teens (fieldwork late 2024). The trend line matters more than the level: every measured population doubles in roughly a year. A policy written for last year's usage rate is already wrong.

3. 60–70% — students admitting at least one cheating behavior per month, for fifteen years. Stanford's Challenge Success surveys. This is the number that reframes everything: cheating was endemic long before ChatGPT. And after ChatGPT? "That percentage has stayed about the same or even decreased slightly in our 2023 surveys." In researcher Victor Lee's words, "the data suggest, perhaps to the surprise of many people, that AI is not increasing the frequency of cheating." Denise Pope adds that when students cheat, "it's typically for reasons that have very little to do with their access to technology."

4. ~3% — submissions Turnitin's own detector scores as mostly (80%+) AI-written. From the vendor's first-anniversary release: of "over 200 million papers" reviewed, "over six million (approximately three percent…) have at least 80 percent AI writing present"; about 11% showed at least 20%. These are unvalidated detector outputs, with error in both directions — but even the detection vendor's maximal reading says the catastrophic case is rare.

5. 53% — students whose main reason for not using AI is fear of being accused of cheating. HEPI again: accusation fear outranks the tools' actual failure modes (hallucinations, 51%) as a deterrent. And 76% of students believe their institution "would spot the use of AI" — a confidence the published accuracy record does not support. Your deterrence is running on fear of a test that misfires most on non-native and formulaic writers.

What the numbers imply for policy

Write policy for the 88%, not the 3%. If nearly nine in ten students use AI on assessed work, a policy whose only categories are "used AI" and "didn't" criminalizes the majority and provides no guidance where students actually need it. Students themselves draw graded lines — in Pew's data, 54% of teens find AI acceptable for researching topics but only 18% for writing essays. A usable policy matches that structure: it names permitted uses, named forbidden uses, and a disclosure habit for the middle. Sample three-tier statements from real syllabi are collected in syllabus AI policy language.

Don't build on the panic premise. The Stanford data removes the foundation under emergency measures: there is no measured cheating surge to suppress. The students who cheat mostly cheated before, for reasons — pressure, workload, disengagement — that predate the technology. That argues for the assignment-design and process-based responses in alternatives to AI detectors over surveillance escalation, and for assignments where the process is the product.

Mind who the fear lands on. The 53% deterred by accusation fear includes your most honest students — the HEPI report quotes one: "I enjoy working with AI as it makes life easier when doing assignments however I do get scared I'll get caught." Fear of false accusation changes behavior in ways that have nothing to do with integrity: students avoid legitimate tools, over-hedge their prose, or pre-emptively confess to things they didn't do in integrity meetings. The conversation guide covers how to keep a concern from becoming a wrongful accusation.

Treat detector flags as base-rate problems. If roughly 3% of papers are mostly AI-written (the vendor's own ceiling) and detectors carry the false-positive rates documented in the independent accuracy literature, then a meaningful share of flags in any given stack are innocent students — concentrated among non-native speakers. A flag is a reason to look at process evidence, never a finding. The evidentiary standards are in what the research says — a briefing for integrity panels.

The constructive move

The numbers describe a gap: students are using AI constantly, mostly within what policies permit, and their dominant anxiety is being unable to prove their use was legitimate. That is a problem evidence solves and detection cannot. A consent-based writing-provenance record — kept by the student as they write, sealed against tampering, and verifiable by anyone in a browser — lets the 88% show which uses happened, and lets you reserve scrutiny for cases where the evidence, not a statistic, warrants it. How to read such a record in a meeting is covered in how to read a process record; Authmarks provides this as an LTI 1.3 tool for Canvas and other LMS platforms, free for individual students.

Go deeper — the primary data, read closely: student AI-use surveys — what the numbers say · AI in scientific writing and journal policies