Student AI-use surveys, read closely — HEPI, Pew, Stanford, and Turnitin's own data

What this review covers

Every policy argument about AI in education eventually reaches for a usage number — "everyone is using it," "hardly anyone is really cheating," "the numbers doubled." This review collects the primary sources those numbers come from and reads them closely: the HEPI/Kortext survey of UK undergraduates (the most-cited higher-education series), Pew Research Center's survey of US teens, Stanford's long-running cheating surveys, and Turnitin's own corpus statistics from 200 million submissions.

Read together, they tell a more precise story than any one headline: AI use is near-universal and still climbing; AI use for assessed work is now the norm; and yet the best longitudinal evidence says overall cheating has not increased. Those three findings are not in tension — they describe a world where the tool is everywhere, the line between legitimate and illegitimate use is blurry, and the population being flagged by detectors is not the same as the population doing misconduct.

Every quotation below is verbatim from the linked source, checked on the asOf date.

HEPI 2025: the near-vertical adoption curve

The Higher Education Policy Institute's Student Generative AI Survey 2025 (authored by Josh Freeman, fieldwork by Savanta, n = 1,041 UK undergraduates) is the cleanest year-over-year measurement in the literature, because it repeated its 2024 questions exactly. The two headline deltas:

"The proportion of students using generative AI tools such as ChatGPT for assessments has jumped from 53% last year to 88% this year."

"The proportion of students reporting using any AI tool has jumped from 66% last year to 92% this year."

In one academic year, not using AI went from a plurality position to a rounding error: "The proportion who have not used generative AI for their assessment in these ways has plummeted from 47% last year to just 12% this year." The most common assessment uses are explanation, summarising articles, and suggesting research ideas — and "'Generating text' is the most popular reason for using AI" overall.

The report's author, Josh Freeman, on the magnitude:

"It is almost unheard of to see changes in behaviour as large as this in just 12 months. The results show the extremely rapid rate of uptake of generative AI chatbots. They are now deeply embedded in higher education and many students see them as a core part of the learning process."

The detail that matters for integrity policy

Buried below the adoption headline are the findings about fear and deterrence. What stops students from using AI is not principle but risk:

"The main factors putting students off using AI are being accused of cheating (said by 53% of respondents) and getting false results or 'hallucinations' (51%)."

And the students believe surveillance works: 76% say "their institution would spot the use of AI in assessments" — a belief the independent detector-accuracy record does not support. One student respondent, quoted in the report:

"I enjoy working with AI as it makes life easier when doing assignments however I do get scared I'll get caught."

That sentence is the student-side experience of the detector regime in miniature: near-universal use, high perceived detection, and fear of accusation — which lands on honest and dishonest users alike, since detectors cannot reliably tell them apart.

The survey also documents a widening equity gap: "There is a growing digital divide in AI use, with male students, students on STEM and Health courses and more socioeconomically advantaged students more likely to use AI than others." And institutions are behind their own students: 67% of students "think using AI is 'essential' in today's world," while "only a third (36%) of students have received training in AI skills from their institution."

Pew 2025: the US school-age picture

Pew Research Center surveyed 1,391 US teens (ages 13–17) from "Sept. 18 to Oct. 10, 2024, through Ipsos." The headline:

"The share of teens who say they use ChatGPT for their schoolwork has risen to 26%, according to a Pew Research Center survey of U.S. teens ages 13 to 17. That's up from 13% in 2023."

Doubling in a year — the same curve HEPI measured, earlier on its slope (US teens, one specific tool, schoolwork only). The more interesting Pew finding is that teens hold a graded view of legitimacy, not a binary one:

• "Just over half of teens say it's acceptable to use ChatGPT to research new topics (54%)."
• "29% of teens say it's acceptable to use ChatGPT to solve math problems, while 28% say it's not acceptable."
• "18% say it's acceptable to use ChatGPT to write essays, and 42% say it's not acceptable."

Students themselves draw the line roughly where most syllabus policies do: research assistance broadly fine, full essay generation mostly not. The policy problem is the vast middle — and a finished-text detector is blind to exactly that distinction, because it only ever sees the final string of words.

Stanford: the cheating baseline nobody expected

Against the adoption curves stands the most under-cited finding in this literature. Stanford Graduate School of Education researchers Victor Lee and Denise Pope run long-term anonymous surveys of US high-school students through the Challenge Success project. The baseline, from Stanford's account: for more than 15 years, "60 to 70 percent of students have reported engaging in at least one 'cheating' behavior during the previous month."

Then ChatGPT arrived, and:

"That percentage has stayed about the same or even decreased slightly in our 2023 surveys."

Lee, verbatim:

"the data suggest, perhaps to the surprise of many people, that AI is not increasing the frequency of cheating."

Pope, on the mechanism:

"we haven't seen that bear out in our data so far. And we know from our research that when students do cheat, it's typically for reasons that have very little to do with their access to technology."

The reconciliation with HEPI and Pew is straightforward once stated: cheating was already endemic (60–70% admitting at least one behavior per month, for fifteen years), and AI largely substituted for older methods rather than minting new cheaters. The students who were going to cut corners found a new tool; the students who weren't, didn't start. Which means the post-ChatGPT integrity panic is substantially a visibility panic — the new method produced a new (and unreliable) way of being caught, not a new population of offenders.

Turnitin's own corpus: what 200 million submissions showed

The fourth data source is not a survey but the largest observational dataset in existence: Turnitin's first-anniversary report on its AI detector (press release, 9 April 2024). Verbatim:

"Since the launch of the tool in April 2023, over 200 million papers have been reviewed."

"Over 22 million (approximately 11 percent of over 200 million) have at least 20 percent AI writing present"

"Over six million (approximately three percent of over 200 million) have at least 80 percent AI writing present"

Taken at face value, Turnitin's own numbers say the catastrophic case — papers that are mostly AI — is about 3% of submissions, and the elevated case about 11%. Two caveats cut in opposite directions. These figures are detector outputs, not ground truth: the same scoring pipeline whose conditional false-positive scoping and sub-20% asterisk policy are documented in vendor claims vs independent measurements, so some of that 11% are false positives — and, per the paraphrase-attack literature, an unknown amount of laundered AI text sits undetected in the other 89%. What the numbers do establish is Turnitin's own view of scale: even by the detector's maximal reading, the overwhelming majority of student submissions are not substantially AI-written.

For global context, the Digital Education Council's Global AI Student Survey 2024 collected "3839 responses across 16 countries, from bachelor, masters, and doctorate students" — its detailed percentages sit behind registration, so this review cites only its public scope, but its existence confirms the survey picture is worldwide, not an Anglosphere artifact.

Summary table

Source	Population	Year-over-year	Headline numbers
HEPI/Kortext 2025 (n=1,041)	UK undergraduates	2024 → 2025	Any AI use 66% → 92%; use for assessments 53% → 88%; 53% deterred mainly by fear of being accused of cheating
Pew Research (n=1,391)	US teens 13–17	2023 → 2024	ChatGPT for schoolwork 13% → 26%; essay-writing acceptable to only 18%
Stanford GSE / Challenge Success	US high-school students	15-year series → 2023	Cheating stable at 60–70% admitting ≥1 behavior/month; post-ChatGPT "stayed about the same or even decreased slightly"
Turnitin corpus (Apr 2024)	200M+ submissions	Apr 2023 → Apr 2024	~11% flagged ≥20% AI writing; ~3% flagged ≥80%
Digital Education Council 2024	3,839 students, 16 countries	—	Global scope confirmation (detail behind registration)

What the numbers do and do not establish

Established:

• Generative-AI use among students is near-universal in the UK sample (92%) and rising fast everywhere measured; use for assessed work is now majority behavior (88% in HEPI 2025).
• Students distinguish sharply between assistance and substitution — only 18% of US teens find essay-writing by ChatGPT acceptable.
• The best longitudinal cheating data shows no increase in overall cheating after ChatGPT; the dominant driver of cheating remains non-technological.
• The primary deterrent students report is fear of being accused — a fear that applies regardless of guilt, given documented detector error rates.
• Even the largest detector vendor's own corpus puts heavily-AI-written submissions at roughly 3%.

Not established:

• That high AI usage equals high misconduct. The surveys measure tool use, most of which (explanation, summarising, research ideas) falls inside most institutions' permitted zone.
• That detector flag rates track real prevalence — Turnitin's 11%/3% are unvalidated detector outputs, bounded by known false-positive and known-evasion error in both directions.
• That students can accurately judge detection risk: 76% believe their institution "would spot" AI use, a confidence the independent accuracy literature does not license in either direction.

Why this matters for provenance

The survey literature describes a population in which nearly everyone uses AI somewhere in their workflow, almost nobody considers all of it cheating, and the main anxiety — for honest students explicitly — is being accused. In that world, a finished-text detector is the worst possible instrument: it cannot see the difference between permitted and forbidden use, it fires disproportionately on non-native and formulaic writers, and it feeds a fear that HEPI's respondents name as their top deterrent — "being accused of cheating," said by 53%, ahead of the tool's actual failure modes.

What the numbers argue for is an instrument that records which uses happened rather than guessing whether text looks machine-made: a consent-based writing-provenance record that shows typing, pasting, and revision as they occurred, sealed so it breaks visibly if tampered with, and checkable by anyone without an account. For the 88% using AI legitimately within policy, it converts "I only used it to summarise sources" from a plea into evidence; for institutions, it replaces a 3%-base-rate guessing game with records — the approach argued across this silo and implemented by Authmarks, free for individual writers.

Related reading: AI in scientific writing and journal policies · how accurate are AI text detectors? · what the research says — a briefing for integrity panels