The Detector Is Not the Judge: A Smarter Approach to AI in UK Coursework

Few topics generate as much anxiety in UK universities at present as the use of generative AI. Alongside that anxiety sits a flourishing market of detection tools, marketed with precise-sounding percentages and colour-coded verdicts. It is worth pausing on an uncomfortable fact: a detector score is not a finding of misconduct, and in most institutions it never has been.

What the software is actually measuring

Detection tools do not identify authorship. They analyse statistical properties of text, such as how predictable each word is given the words before it, and how much variation exists between sentences. Text produced by a language model tends to be smoother and more predictable than human writing, so the software reports a probability.

That is a reasonable signal, but it is not a fingerprint. Any writing that is formulaic, heavily edited or drafted from templates will share some of those characteristics. So will a paragraph that has been run through a grammar checker, translated from another language, or written by someone whose academic English is deliberately plain and careful.

False positives are not rare

Any student who has watched a detector flag a wholly original paragraph will understand the problem intuitively. Because the tools rely on pattern rather than provenance, groups such as international students, mature learners returning to study, and anyone writing in a genre with rigid conventions are disproportionately at risk.

Different products also disagree with one another on identical text, sometimes dramatically. A tool that returns different answers on successive runs is not producing evidence in any meaningful sense; it is producing an estimate. Treating that estimate as a verdict is a category error.

What UK policy generally says

The sector has moved away from blanket bans towards a more nuanced position: what matters is unauthorised use, not the existence of the technology. Many institutions now require students to declare AI assistance where it has been permitted, and to retain responsibility for accuracy, referencing and argument.

Crucially, most misconduct procedures require human academic judgement. A detector report may be considered alongside drafts, supervision records and a conversation with the student about their work. It is seldom decisive on its own, and institutions know this. If you are ever asked about your use of a tool, the question being tested is usually whether the submitted work represents your own understanding.

Three defensible uses

There is a broad consensus that some applications are legitimate when the module allows them. Brainstorming angles for an essay, asking a tool to explain a difficult concept in simpler terms, or requesting structural feedback on a draft you have already written all keep your thinking central.

Less defensible is generating prose that you then submit as your own, fabricating references, or using a tool to paraphrase sources in order to disguise reliance on them. The distinction is not about which button you pressed; it is about whether the intellectual work remained yours.

A practical record-keeping checklist

Transparency is far easier when it is habitual rather than reconstructed after the event.

• Check the specific module and programme policy before you start, since permissions differ between assessments.
• Keep a simple log of prompts, outputs and the date of each session, stored with your notes.
• Retain dated drafts so the development of your argument is visible.
• Include a declaration in the required format if your institution asks for one.
• Cite tools according to your department's referencing style, which will normally follow Cite Them Right conventions.
• Verify every reference a tool suggests against the original source; invented citations are common.
• Avoid entering confidential, personal or unpublished data into public tools.

When to put the tools away

Some moments call for a closed laptop. During the drafting of your analysis, when you cannot explain a sentence you have just written, during timed or invigilated assessments, and whenever the module policy prohibits assistance, the sensible course is to work unaided. If a tool is doing the thinking, the assessment is no longer measuring you, and that will show in a viva, a tutorial question or a subsequent module.

If your work is queried

Respond calmly and factually. Produce your drafts, notes and log if you kept one, and explain your process. Ask what evidence is being relied upon and seek advice from your students' union's academic advice service, which deals with these cases regularly. Honesty about a permitted tool used clumsily is a very different matter from a fabricated denial, and it is usually treated as such.

Integrity, not fear

Detection software is an imperfect instrument being asked to answer a question it cannot fully answer. The more durable protection is the one students have always relied on: work you understand, sources you have read, and a clear account of how the piece came together. Used openly and in proportion, AI can support that process. Used as a substitute for it, it undermines the very thing the qualification is meant to certify.