AI paraphrasing and written assignments: what integrity means now
Paraphrasing someone else's work is still someone else's work
Institutions sometimes fumble the definition here, so state it plainly: taking a source's ideas, evidence, and structure and restating them in fresh words without attribution is plagiarism of the source, whatever tool did the restating. The offence lives in the unattributed borrowing, not in the similarity score. Policies written around percentages of matching text miss their own point, and students exploit the gap in good faith as well as bad, genuinely believing that rewritten means original.
AI tools add a second category: text generated rather than borrowed, which may match nothing and still not be the student's work. That is contract cheating's logic with a faster typist, and policy needs to name it separately, because the evidence, the conversation, and the sanction reasoning all differ.
Why AI detectors cannot carry the case
AI text detectors are probabilistic classifiers with documented false positive rates, and their errors are not evenly distributed: non native English writers are flagged at higher rates by systems that mistake simpler, more regular prose for machine output. A detector score, like a gaze flag, is a reason to look, never a finding. Institutions that have sanctioned on detector output alone have reversed themselves on appeal and damaged trust in every legitimate case heard afterward.
The evidence that decides cases is older and duller: version history showing a document appearing fully formed, drafts that do not exist, style and vocabulary discontinuities with the student's supervised writing, and a conversation in which the student cannot explain choices in their own submitted text. None of it is automatic. All of it is fair, because it is about the work, not a score.
Designing assignments that show their working
Require the artefacts of process as part of the submission: an outline, an annotated source list, one intermediate draft, each lightweight, each hard to fabricate convincingly after the fact. Ask for application to material only the cohort has: the seminar discussion, the local dataset, the text set this term. Add short authorship conversations to a sample of submissions and to any case under question, framed as discussion of the work, which honest students experience as teaching and dishonest ones experience as the end of the road.
None of this requires banning tools. It requires stating, per assignment, which assistance is permitted, declared, or prohibited, and then assessing in a way that makes the declaration checkable. Ambiguity is the policy failure students fall into and cheats hide behind, and it is the one failure entirely within the institution's gift to fix.
Related reading: How do virtual machine tricks let test takers cheat inside locked-down exam browsers?, How do collusion rings coordinate cheating during live online exams?.
Frequently asked questions
If I paraphrase a source in my own words and cite it, is that acceptable?
Yes, that is ordinary scholarship. Paraphrase with attribution is how fields build on each other. The violation is paraphrase without attribution, or paraphrase so close in structure and expression that the citation does not honestly cover how much was taken. When in doubt, quote directly or summarise at a higher level.
Can a university rely on an AI detector score to fail a student?
Institutions that try find the score cannot bear the weight. Detectors are not validated to the standard an academic misconduct finding needs, vendors themselves describe outputs as indicators, and appeal panels ask for evidence about the work. Use a score to start a review. Never as the review's conclusion.
What is the difference between AI assistance and AI authorship?
Where the thinking happened. Assistance with spelling, structure suggestions, or explaining a concept leaves the argument and expression the student's. Submitting generated text as your own work transfers the authorship itself. Policies that specify the line per task, with examples, are enforceable. Policies that say use AI responsibly are not.
How do draft requirements help honest students?
They create the alibi honest students otherwise lack. A student accused wrongly, by a detector or a suspicion, with a trail of drafts and notes can show their process in minutes. Process evidence protects the innocent at least as much as it exposes the dishonest, which is the argument that wins staff rooms.