How Teachers Spot AI-Paraphrased Coursework

The version of AI misuse that gets discussed in staff meetings is the simplest one: a student pastes a question into a chatbot, copies the answer, and submits it. But the version that actually reaches your marking pile is usually a step more sophisticated. A growing number of students now run AI-generated text through a second tool — a paraphraser, a "humaniser," a rewriting app — specifically to soften the patterns that make AI writing recognisable. If you are trying to spot AI-paraphrased coursework fairly, it helps to understand what that extra step does and does not change, and how to keep your process grounded in evidence rather than in a single number.

What Paraphrasing Actually Does to AI Writing

A paraphraser takes a block of text and rewrites it sentence by sentence: swapping vocabulary, reordering clauses, varying sentence length, and breaking up the smooth, evenly-weighted rhythm that fluent AI output tends to have. The marketing promise is that the result will "pass" detection. The reality is more partial. Paraphrasing genuinely does reduce the statistical signal a detection tool relies on, because it disrupts the surface patterns the model learned to associate with AI generation. It does not, however, put the student's own thinking back into the work. The argument, the structure, the choice of what to include — all of that is still the original AI's, just dressed in slightly different words.

That gap is the thing worth holding on to. A paraphrased AI essay can read as oddly generic in a way that survives the rewrite: it covers the obvious points competently, reaches no surprising judgement, and stays curiously detached from the specific texts, lessons, and debates your class actually worked with. The vocabulary may have been shuffled, but the underlying lack of a particular student's voice often remains.

What a Likelihood Score Means on Paraphrased Work

GradeOrbit's AI detection tool returns a likelihood score between 0% and 100% on every piece of work you submit. The score reflects how closely the writing matches the patterns associated with AI-generated text. On paraphrased work, the honest position is that the score becomes less informative, not more. Heavy rewriting can pull a score down from where unedited AI output would sit, which means a paraphrased piece may land in the middle of the range rather than at the top.

This is exactly why the score has never been designed to be a verdict, and why treating it as one is a mistake on paraphrased work above all. A lowered score on a piece that was generated and then rewritten does not clear the work; it simply means the tool's signal has been deliberately muddied. Equally, a mid-range score on a genuine student who writes in a polished register is not evidence of misuse. The number tells you where to look more closely — it does not tell you what you will find. Our guide on handling borderline AI detection scores covers how to weigh a score that sits in that uncertain middle band.

The Signals Paraphrasing Cannot Hide

Because paraphrasing changes the words but not the substance, the most reliable evidence is the same evidence you would use for any suspected misuse: the student's own supervised work. Read the coursework alongside timed essays, class tasks, and assessments the student produced in front of you. The questions are unchanged by paraphrasing. Is the analytical sophistication consistent with what this student produces under your supervision? Does the work engage with the specific sources, interpretations, and terminology your lessons foregrounded, or does it float at a generic level that could have come from any class anywhere?

Paraphrased AI work often carries a few tells of its own. The vocabulary can feel slightly mismatched — unusually elevated word choices sitting next to oddly plain ones, because the rewriting tool substituted synonyms without a sense of register. Transitions can read awkwardly, since reordered clauses do not always reconnect cleanly. And the references, if any, tend to be vague or subtly wrong in a way a student who genuinely read the material would not produce. None of these is proof on its own. Together, alongside a detection score and your knowledge of the student, they build a picture.

Keeping the Process Fair

The risk with paraphrasing in the conversation is over-correction. Once you know students are rewriting AI output to dodge detection, it is tempting to treat every mid-range score as concealment. That is unfair to the many students who write in a clear, formal style and have done nothing wrong. The discipline is the same as ever: the score is one input, your read of the work against supervised writing is a second, and a conversation with the student is the third and most decisive.

That conversation works best when it starts with the content, not the accusation. Ask the student to talk you through a specific part of their argument — why they weighted one point over another, which source led them to a particular judgement. A student who genuinely did the work can reconstruct the thinking; a student who generated and rewrote it usually cannot, because the reasoning was never theirs to begin with. The guide to what to do when a student denies using AI walks through how to hold that conversation when the student pushes back.

Documenting Your Decision

Whatever you conclude, write it down at the time. For paraphrased work especially, a contemporaneous note matters, because the case may hinge on the qualitative evidence rather than the score. A few sentences will do: the likelihood score, why you had concerns, what you asked the student, what they said, and the decision you reached. A record made on the day is far more defensible than one reconstructed weeks later. If your centre's policy does not yet address AI use and the tools students use to disguise it, the school AI academic integrity policy guide is a sensible starting point, and the guide to handling a high likelihood score fairly covers the wider principle of treating the number as evidence rather than judgement.

Running Work Through GradeOrbit

Checking suspected paraphrased work follows the same process as any other detection run. You redact the student's name and identifying details by drawing black boxes over them before uploading — the redaction is burnt into the image in your browser, so the identifying information never leaves your device. For typed coursework you can upload the document or paste the text; for handwritten drafts, scan or photograph the pages clearly. GradeOrbit offers a 1-credit model for quick triage across a set and a 3-credit model for higher-confidence analysis on a piece that matters more. Student work is never stored — it is processed and discarded once the score is returned, so there is no record on any server of which student was flagged or what their work contained.

Try GradeOrbit for AI Detection

If you want an AI detection tool that gives you a clear likelihood score to weigh alongside your own professional judgment — including on work that students have tried to disguise — GradeOrbit is built for exactly that. New accounts get a small allocation of free credits to try the detection workflow on real student work before any commitment.

Visit gradeorbit.co.uk to learn more and get started. The tool takes minutes to set up and works on the first piece of work you upload.