Student AI Use in 2026: What the Numbers Actually Measure

The headline finding: 94% of surveyed students say they use generative AI to help with assessed work, and 12% say they directly include AI-generated text in it. Both figures come from the same survey. Nearly every strategic misreading in this area comes from treating the first number as if it meant the second.

Below, the figures with their sources, samples and dates attached — and an explicit account of what they cannot tell you.

The source and its sample

The most consistent published series is HEPI’s Student Generative AI Survey. The 2026 edition is HEPI Report 199, by Rose Stephenson and Charlotte Armstrong, published 12 March 2026 and sponsored by Kortext. It was conducted by Savanta in December 2025 and is based on responses from 1,054 full-time UK undergraduates.

The 2025 edition is HEPI Policy Note 61, by Josh Freeman, published 26 February 2025, based on 1,041 online interviews conducted by Savanta in December 2024, weighted on demographics including gender, institution type and year of study, with a margin of error of approximately 3%.

State that sample every time you quote the figures, because it constrains them severely. These are UK respondents, full-time students, and undergraduates. They are not postgraduates, not part-time students, and not a cross-country sample. If your business case concerns doctoral supervision, this survey does not measure your population.

It is also worth noting who commissions this work. Both editions are sponsored by a commercial partner, which is disclosed on the reports and is entirely normal for sector research. It does not invalidate the findings — the fieldwork is run by an independent survey house and the data tables are published — but it is a fact to carry alongside the numbers rather than to discover later in a committee.

The two numbers, and why the gap matters

The 94% figure encompasses activities most institutional policies permit outright, and several that programmes now actively teach. It is a measure of tool adoption. The 12% figure is the one that bears on authorship, and even it is not a misconduct rate — inclusion of AI-generated text is permitted in some assessments, and where permitted with disclosure it is compliance rather than breach.

So there is no published figure here for “students cheating with AI,” and any presentation that offers one is doing arithmetic the survey does not support. What the series does give you is a trend on the narrow measure, and it is a clean one because HEPI reports it as a series: 3% in 2024, 8% in 2025, 12% in 2026.

That trajectory — roughly a quadrupling over two years from a very low base, on a measure that still describes a small minority — is a better basis for planning than either the alarm or the reassurance usually attached to this topic.

Why adoption estimates vary so much between studies

Institutional researchers regularly encounter two credible-looking studies reporting AI adoption twenty points apart, and conclude that one must be wrong. Usually neither is. Four design choices move the number substantially before any real behaviour is involved.

• The verb. “Have you ever used” produces a far higher figure than “do you regularly use,” which in turn exceeds “did you use, for this assessment.”
• The object. “AI tools” includes spelling and grammar checkers many students do not think of as AI; “generative AI” is narrower; “ChatGPT” narrower still.
• The population. Undergraduate coursework, taught postgraduate work and doctoral research have different task structures, so a single adoption rate across them is an average of unlike things.
• The framing. A question preceded by academic integrity language depresses reported use, because it signals that an answer may be judged.

The practical consequence is that a difference between two studies is usually a difference between two questions. Before reconciling figures, put the two question wordings side by side; if they differ, there is nothing to reconcile.

What students say about their institutions

The institutional-performance figures are the ones most useful for a business case, because they describe a gap the institution controls.

Three readings worth taking from that table.

The population is split, not aligned. In 2026, 37% agree their institution encourages AI use and 36% disagree. A near-even split among students at the same institutions is a communication finding, not an attitude finding: it indicates that what is permitted is not reliably reaching the people doing the assessments.

Provision lags expectation by 30 points. 68% believe AI skills are essential; 38% say their institution provides tools. Whatever else that gap describes, it describes students buying capability privately, which turns attainment into a function of household budget.

Staff capability is improving from a very low base. The proportion saying staff are well-equipped moved from 18% to 42% between the 2024 and 2025 waves. That is real progress and still leaves a majority unconvinced.

The figures that get misused

Four specific cautions, each of which we have seen produce a wrong decision.

1. Do not convert 94% into a misconduct estimate. It counts permitted activity. The narrow measure is 12%, and even that is not a breach rate.
2. Do not apply UK undergraduate figures to postgraduate populations. The survey does not sample them, and thesis-stage work has a different task structure entirely.
3. Do not compare across surveys with different wording. Adoption estimates vary widely by instrument, which is why the internal HEPI series is usable and cross-vendor comparisons generally are not.
4. Do not treat self-report as behaviour. These are students describing their own conduct in a context where some of it may be prohibited. Self-report on sensitive behaviour is conventionally understood to under-report.

Point four cuts against our own framing, which is why it belongs here: the 12% is plausibly a floor rather than a central estimate. Presenting it as precise would be as wrong as presenting 94% as a cheating rate.

What we could not verify

In preparing this piece we attempted to retrieve several other frequently cited sector sources directly and were blocked by automated challenges rather than served content. We have therefore excluded them entirely rather than cite figures at second hand.

That exclusion is itself informative for institutional researchers: a good deal of what circulates as “sector data” on AI adoption cannot be traced by a reader to a document with a stated sample and method. If a figure in your board paper cannot be traced to an opened source, it should not be in your board paper.

How to use these numbers in an institutional paper

1. Quote the pair, never the single figure. “94% use generative AI to help with assessed work; 12% include AI-generated text directly” is honest and instantly reframes the discussion.
2. Carry the sample. Survey house, month of fieldwork, n, and population, in the sentence or the footnote.
3. Use the trend, not the level, for planning. 3 → 8 → 12 tells you what to prepare for; a single year’s level tells you where you were.
4. Pair national figures with a local instrument. A short, stable, repeated internal survey is worth more for your decisions than any national number, because it samples your students.
5. Fix the wording and never change it. The only reason the national series is usable is that its questions stayed comparable.

That fourth point is the one institutions most often skip and most often need. National data tells you the direction of travel; it cannot tell you whether your policy is landing. The measurement design for a local instrument is covered in our guide to running a departmental pilot, and the policy components those numbers should inform are in what a university AI policy should include.

If you would like to discuss what a defensible internal measurement baseline looks like alongside a platform evaluation, request an institutional evaluation.

Frequently asked questions

What percentage of students use AI?

95% reported using AI in at least one way in HEPI’s 2026 survey of 1,054 full-time UK undergraduates.

How many use it for assessed work?

94% in the same survey — a measure that includes explaining concepts, summarising and checking phrasing.

How many submit AI-generated text?

12%, up from 8% in 2025 and 3% in 2024.

Is 12% a cheating rate?

No. Including AI-generated text is permitted in some assessments, and where it is permitted and disclosed it is not misconduct.

Do these figures cover postgraduates?

No. The sample is full-time undergraduates in the UK.

Are the figures internationally comparable?

Not reliably. Estimates vary substantially by instrument and population, so cross-survey comparison should be avoided unless the wording matches.

How reliable is self-reported data on this?

Treat it as a floor. Self-report on potentially sanctionable behaviour conventionally under-reports.

Do students think their institutions support them?

They are split: 37% agree their institution encourages AI use and 36% disagree, while 48% feel teaching staff help them develop AI skills.

How many students get AI tools from their institution?

38% say they are provided with AI tools.

Has assessment changed?

65% of students in 2026 say assessment has changed significantly in response to AI.

Why is my figure different from another report’s?

Almost always sample and wording. Before reconciling two numbers, check whether they measure the same behaviour in the same population — the general problem covered in our piece on how current your higher education data actually is.