How to be honest about using AI

Most students think the question is "am I allowed?" The real question is whether you could say what you did, out loud, to your tutor, without leaving bits out. If you could not, you already know the answer, and no policy document is going to change it. The useful test is not whether a machine touched the work. It is whose judgement is in it.

Four things that make the answer easy#

Say so before you are asked. More and more departments now penalise not declaring rather than using. Declaring costs nothing and reads as somebody who knows what they are doing.

Keep the receipts. Write in a document with version history switched on. Keep your notes and the ideas you dropped. That record shows how you thought, which is a different and better thing than what you handed in.

Log it as you go. Four columns: the task, what the AI did, what you did, how you verified it. Two minutes per assignment, written while you still remember.

Apply the out-loud test. Could you describe the whole process to your tutor without flinching? That is the answer, and you already have it before any policy is consulted.

Why "will I be caught" is the wrong question to organise around#

Detection is not a reliable adversary in either direction, and building a strategy on it fails twice over.

At the University of Reading, 100 fabricated submissions written entirely by AI were entered into five real undergraduate modules through the live marking system. 94 per cent went undetected, and on the stricter test of a marker actually mentioning AI, 97 per cent. They also outscored real students by just over half a classification boundary. So the deterrent is weaker than most students assume.

And it is unreliable in the other direction, which is the half nobody mentions. Detectors misclassified more than half of essays by non-native English writers as AI-generated, an average false positive rate of 61.22 per cent, while classifying US eighth-grade essays almost perfectly. The proposed mechanism is that detectors key on predictability, and second-language writing is more predictable. If you write in your second language, a clean process and a written record are worth more to you than to anyone else in the room.

Meanwhile the penalties are real and rising. Russell Group universities recorded 2,053 punishments in 2024-25 against roughly 700 the year before, with four members disclosing expulsions. Seven of the twenty-four do not record AI investigations at all, so that is a floor on an incomplete sample rather than a national picture.

Declaring is more normal than the anxiety suggests#

The HEPI student survey is the number worth knowing. It is also routinely misquoted. 94 per cent of UK undergraduates say they have used generative AI to help prepare assessed work. That is mostly comprehension: explaining concepts 61 per cent, summarising an article 49, suggesting research ideas 40, structuring thoughts 39.

Including AI-generated text directly in assessed work is 12 per cent. The 94 and the 12 are answers to different questions, and a student who reads the 94 as "everybody is doing what I am worried about doing" has misread the survey they are taking comfort from.

What a declaration actually looks like#

Not a confession, and not a disclaimer nobody reads. Four lines, specific:

"I used Claude to explain two passages in the Hobsbawm chapter that I could not follow. I used it to generate five counter-arguments to my thesis, of which I used two. The reading list came from the library catalogue and I opened every source. All prose is mine. I did not use it to draft or edit any sentence in this essay."

That takes two minutes, it is checkable, and it describes a process a marker would be pleased to see. A student who cannot write four lines like those has learned something useful about their own process.

The check that survives every policy change#

University rules differ by department and are being rewritten every year, so a rule learned this term may not hold next. The underlying question does not change, and this research asks organisations the same one: what stays mine.

If the tool touched your understanding, you are almost certainly fine and probably better off. If it touched the words that get marked, declare it or do not do it. And if you cannot say which of those happened, that is the finding.

What this page cannot tell you#

It cannot tell you your department's rule, and no page can: official guidance differs between institutions and often between modules in the same institution. Ask your module lead by email, so the answer exists in writing.

Nor is there any evidence that declaring improves marks, or that the four-column log reduces misconduct findings. Nobody has measured either. The case for them is that they cost two minutes and make an unanswerable question answerable, which is an argument from structure rather than a result.

Key sources

• Scarfe, P. et al. (2024). A real-world test of artificial intelligence infiltration of a university examinations system.
• Liang, W. et al. (2023). GPT detectors are biased against non-native English writers.
• Higher Education Policy Institute (2026). Student Generative AI Survey 2026.
• Recorded penalties for AI misuse in UK universities, compiled review.

Related SuperSkills research#

For students, how to use AI at university, what to do when your department bans it and group work when everyone has AI. On detection, does AI detection work. On the decision underneath, keep, share, hand over. On the same question at work, proving you did the work.

About this research#

The out-loud test and the four-column log are used by Rahim Hirji in teaching and appear in the Mastering AI deck, most recently in September 2026. No claim of first use is made for either. Findings are attributed to the studies that produced them and kept separate from the interpretation, which is that disclosure is a better organising question than detection.