How to use AI at university

There is a version of this advice that says use less AI, and a version that says use more, and both are written by people who will not be sitting your exams. This is for the student who is going to use it anyway. The question worth answering is not whether to, but which parts of the work must stay yours if you want to leave with something more than a certificate.

Why "will I get caught" is the weakest reason available#

In summer 2023 researchers at the University of Reading created 33 fake student accounts and submitted answers written entirely by GPT-4 into the live examinations system of a real psychology school, across five undergraduate modules. Markers were staff and trained postgraduates, marking anonymously, and none of them knew the study was happening.

94 per cent of the AI submissions were never detected. On the stricter test, where a marker had to actually mention AI rather than flag anything at all, it was 97 per cent. And across the five modules there was an 83.4 per cent probability that the AI answers would outscore an equal-sized random draw of real students, by just over half a classification boundary. Graded entry.

The authors are careful in a way the coverage usually is not. They say plainly that they cannot estimate how many real students in that cohort used AI, and that their own six per cent detection rate "likely overestimates our ability to detect real-world use of AI to cheat in exams", because a real student would not take an approach as naively obvious as theirs. The 83.4 per cent is a resampling probability against a median, not a count of head-to-head contests won, which is how it usually gets repeated.

Both things are true at once, and the combination is uncomfortable in both directions. Thousands of students are being penalised each year. Most people who do this are never caught. So if your only reason for not doing it is that you might be, you are resting on the one argument the evidence does not support.

What is actually happening to people#

Three separate Freedom of Information investigations, covering different institutions and different years, and they should not be added together. The Times found 2,053 recorded punishments across Russell Group universities in 2024-25, against roughly 700 the year before, with four members disclosing expulsions. The Student Eye found Bristol issued 526 penalties in 2023-24, against seven two years earlier. The Scotsman found 1,051 cases across Scotland in 2023-24 against 131 the year before, of which Abertay alone recorded 351.

Read those numbers with the caveat the universities themselves give. Seven of the 24 Russell Group members do not record AI investigations at all, so 2,053 is a floor across an incomplete sample. Abertay created "unacceptable AI use" as a category only in 2023 and accounts for a third of the Scottish total. Much of the rise is a new box on a form rather than a new behaviour, and anyone quoting these figures as a measure of student honesty is quoting them wrongly. Graded entry.

What is not in doubt is the shape of the consequence. Expulsion is rare and sits at the top of a tiered scale. What usually happens is smaller and worse: a zero on the assignment, or the module failed, settled informally without a hearing.

The bibliography problem, and why it is not about to be fixed#

An audit of 2.5 million papers in the PubMed Central Open Access collection found that the share carrying at least one reference to a study that does not exist rose from one in 2,828 in 2023, to one in 458 in 2025, to one in 277 in the first seven weeks of 2026. Review articles ran 57 per cent higher than other paper types. At the time of the audit, 98.4 per cent of the affected papers had received no publisher action. Graded entry.

These are researchers. People whose whole training is checking sources, publishing in journals with editors and reviewers between them and print. The same pattern reaches courtrooms: a database of decisions where a judge found somebody had relied on hallucinated material passed two thousand entries across 42 jurisdictions by September 2026, and more than eight hundred of them involved practising lawyers. Graded entry.

The rule that follows is short. Never cite a source you have not opened. Use the machine to find the search terms, the debates and the likely authors. Then open the original yourself, read the abstract, method, findings and limitations, record the reference properly, and cite the original rather than the summary. The failure is not that the model lies to you. It is that a plausible reference costs nothing to generate and twenty minutes to check, and at two in the morning nobody checks.

The number everyone quotes, and the one that matters#

HEPI's 2026 survey of 1,054 full-time UK undergraduates found that 94 per cent had used generative AI to help prepare assessed work, up from 89 per cent in 2025 and about 53 per cent in 2024. That figure travels well and is almost always restated as students using AI on their assessed work, which is a different claim and a much larger one.

What the 94 actually covers is mostly comprehension. Explaining concepts, 61 per cent. Summarising an article, 49. Suggesting research ideas, 40. Structuring thoughts, 39. The figure for including AI-generated text directly in assessed work is 12 per cent, up from 3 in 2024. So one preposition turns a 12 per cent finding into a 94 per cent one, and most of the coverage makes exactly that swap. Graded entry.

The useful thing in that data sits below the headline. 68 per cent of students think AI skills are essential, while only 36 per cent feel their institution encourages them to use it. Everyone already has access, so the shortfall is instruction: almost nobody has been taught what the thing is for.

Four levels, and almost nobody leaves the first#

Where a student sits on this is more predictive than which tool they use.

• Extract. "Give me the answer." Faster output, weaker learning.
• Explore. "Help me understand this." You actually understand it afterwards.
• Examine. "Tell me where I am wrong." You get better at spotting what is true.
• Extend. "Help me build something I could not build alone." You can do things you could not do before.

Almost everybody stays on the first, including people who use it every day. The move from one to two costs nothing and takes about a fortnight of deliberate awkwardness. It is the whole difference between a degree that made you capable and a degree that documented your attendance.

The habit underneath all of it#

Think, then AI, then think. Write your own position first, even one line, so you know what you actually believe before the machine tells you what to believe. Then bring it in, and ask it to argue against you rather than agree. Then come back and decide what to keep, what to reject, and what you could defend out loud with the laptop shut.

Almost everyone does the middle part. Hardly anyone does the first and the last, and those are the two that do the work. The version of this that survives contact with a deadline is simpler still: if you cannot defend it without the machine open, you have not learned it, whatever the mark says.

Nobody has followed a student through three years of this#

That AI use lowers grades. The Reading study found the opposite in the short run, and that is the point rather than a complication: the mark and the capability come apart, and the mark is the thing you can see. There is no study measuring what happens to a student who spends three years at level one, because the people who would be in it are still at university. The argument here is a mechanism with strong support from adjacent evidence on how humans learn with AI and deskilling, and no direct test yet.

It also does not tell you what your department allows. That varies by institution, by faculty and sometimes by module, it has changed for your year specifically, and the brief is the only authority on it. Read the brief.

Key sources

• Scarfe, P., Watcham, K., Clarke, A. and Roesch, E. (2024). A real-world test of artificial intelligence infiltration of a university examinations system. PLOS ONE, 19(6), e0305354. Graded entry.
• Topaz, M., Roguin, N., Gupta, P., Zhang, Z. and Peltonen, L.-M. (2026). Fabricated citations: an audit across 2.5 million biomedical papers. The Lancet, 407, 1779-1781. Graded entry.
• Charlotin, D. AI Hallucination Cases. HEC Paris. A live tracker; figures here were read on 5 September 2026. Graded entry.
• Gibbins, A. (2025). University of Bristol sees 7,414% rise in AI penalties. The Student Eye, 24 May 2025, from a Freedom of Information request. Graded entry.
• Stephenson, R. and Armstrong, C. (2026). Student Generative AI Survey 2026. HEPI Report 199, with Kortext. 1,054 full-time undergraduates. Graded entry.
• Mollick, E. and Mollick, L. (2023). Assigning AI: Seven Approaches for Students, with Prompts. Wharton School Research Paper. The authors describe their approaches as "largely untested". Graded entry.
• Ross, C. (2025). Scottish universities catch students misusing AI in more than 1,000 cheating cases. The Scotsman, 1 March 2025, from FOI data obtained by Miles Briggs MSP.