Ninety-five percent of full-time undergraduates surveyed by the Higher Education Policy Institute in December 2025 reported using AI in at least one way, and 94% used generative AI to help with assessed work. In the same system, 5.1 students per 1,000 were formally proven to have cheated with AI in 2023-24. Both figures describe the same population.
The gap between them is the whole question. Using AI is not misconduct. Misrepresenting who did the thinking is.
Sources: HEPI/Kortext Student Generative AI Survey 2026 (Savanta, December 2025); Guardian Freedom of Information survey of 131 UK universities, June 2025; Liang et al., Patterns, 2023.
Is it cheating if you use AI?
Not by default. Almost no university policy bans AI as a category of software. What policies ban is submitting work that misrepresents your own authorship, and AI is one of several ways to do that.
Read your institution's academic integrity code closely and the operative words are contract cheating, misrepresentation, or unauthorised assistance. The offence is defined by the claim you make when you submit, not by the tool that produced the text.
This matters because it means the same prompt can be legitimate in one course and expellable in another. A grammar pass on your own draft is permitted in most writing units. The identical grammar pass is misconduct in a unit assessing your written expression, because expression is the thing being graded.
The 2025 HEPI survey found 18% of students had inserted AI-generated text directly into submitted work. That is the behaviour policies target. It is not the same as the 94% who used AI somewhere in the process.
What do universities actually permit?
The clearest published framework is the AI Assessment Scale, developed by Perkins, Furze, Roe and MacVaugh in 2024 and now promoted by Australia's Tertiary Education Quality and Standards Agency as an assessment transparency tool. It defines five levels, and the authors are explicit that no level is inherently better than another.
The practical value is that your assignment brief probably maps to one of them, even if it never names the scale.
Level 2 is the trap. Students read "you may use AI to help plan" as permission to let the tool write the paragraph they planned. The scale does not say that.
If your brief is silent, assume Level 2 and ask. Silence is not permission, and it is not prohibition either. It is an unanswered question that only the unit coordinator can close.
How to Tell if It's helping or Cheating?
Three tests settle almost every case. Run them in order, and stop at the first failure.
The authorship test. Could you defend every claim, structure choice and source in a five-minute oral exam, with the document closed? If a sentence exists in your draft that you could not have written and cannot explain, it is not yours.
The disclosure test. Would you be comfortable writing down exactly what you did in an AI-use statement attached to the submission? Hesitation here is diagnostic. Misconduct is almost always the thing you would rather not put in writing.
The learning-loss test. After the AI finished, do you know more than before? Or did the task complete while you stayed still? This one has no disciplinary force, but it is the test that predicts your exam mark.
The fourth row is the one students misjudge most. Feeding a source paragraph into a paraphraser feels like the opposite of plagiarism, because the words change. The words were never the point. The idea is still someone else's, and it is still uncited.
If paraphrasing is the actual skill you need, our guide to paraphrasing without losing the citation covers the mechanics, and the essay structure walkthrough covers where paraphrase belongs in an argument.
How many students actually get caught?
The Guardian obtained Freedom of Information data from 131 UK universities in June 2025. Proven AI misconduct cases tripled in one year.
Conventional plagiarism moved the other way, falling from 19 cases per 1,000 students in 2022-23 to 15.2 in 2023-24. Students did not become more honest. The method changed.
Two caveats keep this number honest. More than 27% of responding universities did not record AI misuse as a separate category at all, so the true count is higher. And a 2024 University of Reading study found that assessment systems failed to detect AI-generated submissions 94% of the time.
Read together, those facts point one way. Detection is weak, enforcement is rising fast, and the enforcement curve is steeper than the detection curve. Betting on the 94% is betting against a moving target.
What is the 30% rule for AI?
It is not a rule. Turnitin has never published a 30% threshold, and no major institution defines misconduct by that number.
The phrase circulates in three unrelated senses. In student forums it means "under 30% on a detector and you are safe." In business writing it means AI does 70% of routine work while humans keep 30% for judgment. Harvard Business School's Tsedal Neeley uses it for AI literacy, arguing you need roughly 30% of the core concepts to use these tools well.
Only the first version is about academic integrity, and it is the one that is false. Turnitin explicitly states that it does not set a disciplinary cutoff. Each institution decides.
The number persists because it feels like the middle. A 6% score reads as noise, an 85% score reads as a confession, and 30% sits in the zone where suspicion starts. That is a description of instructor psychology, not policy.
Is a 40% AI score bad for AI Detection?
A 40% score means the panel will look. It does not mean the panel will find. The distinction is load-bearing, and Turnitin's own documentation supports it.
The false-positive problem is not evenly distributed. A Stanford study published in Patterns in 2023 found seven detectors classified 61.22% of TOEFL essays by non-native English writers as AI-generated, while scoring near-perfectly on native writing.
Vanderbilt University disabled Turnitin's AI detection in August 2023 after calculating that a 1% false-positive rate across roughly 75,000 papers would produce about 750 wrongful accusations a year. Not every institution has done the arithmetic.
The defence against a wrong flag is not a lower score. It is a process trail: version history, dated notes, annotated sources, a draft that visibly evolved. AskSia's AI detector gives sentence-level flags on your own drafts before submission, which turns an ambush into a known quantity. It is a diagnostic, not a laundering tool.
Why does using AI feel like cheating?
Because the effort disappeared and the output did not. Every prior study tool left a residue of labour. AI does not, and the absence registers as guilt before your judgment catches up.
The feeling is often wrong. It is also sometimes the only alarm you get, and it fires most reliably when the learning-loss test has failed.
The institutional picture reinforces the confusion. In the 2026 HEPI survey, 37% of students said their institution encourages AI use and 36% said it does not. Sixty-eight percent believe AI skills are essential; only 48% think staff are helping them build those skills.
Students are being told simultaneously that the tool is their future and their downfall. That is not a moral failure on their part. It is a policy vacuum, and it is discussed in more depth in our overview of AI in education and our review of the current AI study tool landscape.
Frequently Asked Questions
Is it cheating if you use AI?
Only if you misrepresent authorship. University codes define misconduct as unauthorised assistance or misrepresentation, not as tool use, which is why 94% of UK undergraduates used generative AI for assessed work in 2026 while only 5.1 per 1,000 faced proven misconduct findings. Using AI to explain a concept is equivalent to using a textbook. Using AI to produce sentences you then submit as your own is misconduct at AI Assessment Scale levels 1 through 3, which covers most undergraduate essay briefs. The test that settles it: could you defend every claim in a five-minute oral exam with the document closed? Check your unit's assessment brief for a stated AI level, and if it is silent, email the coordinator rather than assuming permission.
What is the 30% rule for AI?
There is no 30% rule. Turnitin has never published a 30% threshold and states explicitly that it does not define a disciplinary cutoff, leaving that to each institution. The number originated in student forums as a rumour and hardened into folklore. Separately, business writing uses a 30% rule to mean AI handles 70% of routine work while humans retain 30% for judgment, and Harvard Business School's Tsedal Neeley uses it to describe the share of AI concepts you need for basic literacy. None of these is a safe harbour. Turnitin suppresses all scores between 1% and 19%, displaying them as an asterisk, so a score under 30% carries no defensive weight at all. Find your institution's actual stated threshold and rely on that instead.
Is 40% AI detection bad?
It triggers review at most institutions, but it does not prove anything on its own. Turnitin's under-1% false-positive claim applies only to documents scoring above 20%, and the company states its scores carry roughly ±15 points of variance, so a reported 50% may reflect anywhere from 35% to 65% actual AI content. A 40% result therefore spans a wide band. The risk is uneven: a Stanford study found seven detectors misclassified 61.22% of TOEFL essays by non-native English writers as AI-generated. Vanderbilt disabled the feature entirely for this reason. If flagged, produce your version history, drafts and annotated sources. A process trail beats a score argument every time.
Can professors tell if you used AI?
Systematically, not well. A 2024 University of Reading study found assessment systems failed to detect AI-generated submissions 94% of the time. Detection improves sharply when a human reads for substance rather than style: fabricated citations, sources that do not say what the essay claims, arguments that dissolve under a single follow-up question. Over 27% of the UK universities in the Guardian's 2025 survey did not even track AI misuse as a separate category, which means low reported rates often reflect low measurement, not low incidence. Enforcement is nonetheless tripling year on year. Use AskSia's Concept Map to check whether you can reconstruct the argument without the tool open, which is the same test an oral defence applies.
What happens if you get accused of AI misconduct?
Outcomes range from a mark of zero on the assessment to suspension, and the penalty usually turns on intent and prior record rather than on the detector score. Nearly 7,000 proven cases were recorded across 131 UK universities in 2023-24, with figures projected to reach 7.5 per 1,000 students. Most institutions run a two-stage process: an informal conversation with the marker, then a panel if unresolved. Evidence of a humaniser or paraphrasing tool is treated as an aggravating factor almost everywhere. Bring version history, timestamped drafts, browser or library records, and your notes. Request the specific policy clause you are alleged to have breached, and ask what evidence beyond the score exists.
How should you use AI without cheating?
Keep the AI upstream of the writing and downstream of the thinking. Use it to explain what you did not understand, to interrogate a reading, to quiz you before an exam, and to check your own draft, never to produce prose you then submit. Attach your sources so answers cite the passage rather than inventing one, since fabricated citations account for a large share of detected cases. Disclose whatever your unit asks you to disclose, and write it down before you are asked. The AI ethics workspace and the essay writing workspace both keep the source trail attached to the answer.
When Using AI is Cheating?
Three places, and they are worth naming.
The hardest numbers here are British. HEPI surveys UK undergraduates and the Guardian's FOI data covers UK institutions. US, Australian and New Zealand enforcement rates are less systematically published, though the AI Assessment Scale is promoted by Australia's TEQSA and the underlying behaviour appears comparable.
The authorship test also assumes you have a policy to comply with. If your unit brief says nothing, the test tells you what is defensible, not what is permitted. Those are not the same standard, and only your coordinator can close the gap.
Finally, none of this addresses the quiet cost. A submission can pass every integrity test and still teach you nothing, and the exam hall does not accept a process trail. Pair whatever you use with retrieval practice from our guide to studying effectively, because the authorship test and the exam are ultimately asking the same question.
