It is a question that comes up more and more often in our conversations with institutions. Now that students can have a language model write the assignment for them, is there any point in still checking submissions for similarity? If the copying has moved somewhere a similarity service cannot follow, why keep paying for one?
Behind the question sits a phrase that has travelled a long way since 2023: the post-plagiarism era. It is a useful phrase, and it describes something real. It is also being used to argue for something its author never argued for – that institutions can now stop checking.
They cannot. Not because the technology of cheating has stood still, but because almost everything around it has.
The term comes from Sarah Elaine Eaton at the University of Calgary, whose Six Tenets of Postplagiarism (2023) has become one of the most cited short pieces in the field. Her tenets are worth reading properly, because they are more careful than the shorthand suggests: hybrid human-AI writing will become normal; human creativity is enhanced; language barriers disappear; humans can relinquish control, but not responsibility; attribution remains important; historical definitions of plagiarism no longer apply.
Read that list again and notice what is not on it. Eaton does not say plagiarism has ended. She does not say students have stopped copying. She says nothing at all about whether institutions should stop checking. Two of the six tenets point firmly the other way: responsibility stays with the human, and attribution still matters. The argument is that our definitions need to grow up, not that our practice can be retired.
The version circulating in sector debate is a flattened one – post-plagiarism as “plagiarism is over, so the plagiarism tool can go”. That is a misreading, and an expensive one, because the institution that acts on it will find out slowly and publicly that it was wrong.
If generative AI had displaced copying, the misconduct statistics would show it. They do not.
Our white paper on historic trends in academic misconduct tracked reported cases across the Nordics and the UK. Plagiarism accounts for around 60% of disciplinary cases in Sweden. Swedish cases rose 73.6% between 2015 and 2019, from 880 to 1,528 – before anyone had heard of ChatGPT – and then nearly doubled again by 2021. Norway rose 132.5% over the same pandemic period, Denmark 74.6%. Volumes have come down from the 2021 peak, but they remain clearly above where they sat before 2020.
The trend line matters more than any single figure. The rise began years before generative AI arrived, which tells you the drivers are structural: more students, more assessment at distance, more submissions, more pressure.
Those published figures stop short of the period everyone is now arguing about. Generative AI only reached students at scale in 2024 and 2025, and national misconduct statistics run several years behind – so nobody yet has a clean, public dataset for the AI years. What we can say is what we see, and what institutions tell us. Across the universities we serve, there is no marked fall in the number of plagiarism cases.
The most likely reading is that volumes have settled back towards something like the pre-Covid norm and stayed there – and the reason they have stayed there is not that students ignored the new tools. It is that AI did not replace the old method. It added a variant of it.
Copying a passage outright used to mean hoping nobody noticed. Now the student pastes it into a model and asks for a rewrite. The source is the same, the argument is the same, the structure is the same. Only the wording is new. Washing a text was always possible, but it took time and some skill. It now takes seconds, and anyone can do it – which turns a niche technique into a default one.
So the method has broadened rather than moved. Verbatim copying continues. Machine-generated work has been added at one end. And in between sits a large and growing middle: real sources, laundered wording. That middle is a substantial part of why the case numbers have not fallen – and why “AI has replaced plagiarism” does not match what assessors are actually finding in front of them.
That is the first answer to the question. You do not stop checking for something you are still finding.
The second answer is about where the copying comes from, and it is the part of the picture most often skipped.
Across the submissions we see, the dominant pattern is not machine-generated text. It is peer-to-peer: one student’s work appearing inside another’s, across a cohort, across a campus, and across institutions. We set this out in Heads We Win, Tails You Lose, our report on why AI detection cannot carry an institution’s integrity case – and the finding that surprises people most in that report is not about AI at all. It is that the old problem is still the big one.
The reason is worth saying out loud. Copying from another student is easy. The text already exists, it was written for this course, this brief, this marker’s expectations. Getting a model to produce something that hits the mark – that answers the actual question, at the right level, with the right reading behind it, and sounds like a person who attended the module – takes prompting skill, subject knowledge and several rounds of editing. It is real work. A friend’s assignment from last year requires none of it.
So the route of least resistance has not moved. It still runs through other students’ work – and other students’ work is precisely what a similarity service, checking across cohorts, across years and across institutions, is built to find.
The middle ground described above is where the detection question is actually decided, and it is exactly what the older generation of text-matching tools was worst at. A laundered passage offers nothing for string matching to find: the sentence it was looking for no longer exists. Meaning-level comparison does find it, because the meaning is the one thing the student could not afford to change.
That is the design argument for semantic similarity rather than string matching. “AI has made copying undetectable” has it backwards: AI has made one category of copying harder to see with the tools of 2015, and perfectly ordinary to see with the tools of today. Choosing not to check because the copying has been rewritten means choosing not to look at the fastest-growing part of the problem. We go through the practical differences between the approaches in our deep dive on choosing a plagiarism tool.
Similarity checking does not only catch. It deters, and the deterrent is not a secret – students know their work is compared, and they largely know what that means.
The research is blunt about this. A systematic review by Hattingh, Buitendag and Lall, synthesising 37 studies, ranked the most common reasons students plagiarise deliberately. Lack of deterrence – the judgement that the chance of being caught is small, or the penalty light – sits among the top five. Students in that group are explicitly weighing risk against reward, and some of them already assume that rephrasing or translating will get them past a checker. A stricter policy does nothing on its own; it is only the probability of detection that moves the calculation.
Institutions that quietly stop checking should assume the change is public within one semester. Cohorts talk. Reddit talks. The message that travels is not the nuanced one about assessment redesign; it is “they don’t check any more”. And the behaviour that follows is not sophisticated AI misuse. It is the oldest and simplest thing in the book: a file passed between two students who share a flat.
This is the part of the post-plagiarism argument that has no answer. Remove the check and you do not arrive in a world of authentic hybrid writing. You arrive in a world where copying is cheap, unmonitored, and known to be unmonitored.
The strongest version of the sceptical case is not really about detection at all. It is about design: stop policing the artefact, redesign the assessment, and the problem takes care of itself.
We agree with the destination. We have written about assessment formats that hold up better under AI, about oral defences, about layered exam security and about why the exam hall is not the only answer. None of it happens quickly.
Assessment formats sit inside programme specifications, regulations, external examiner arrangements, accreditation and student contracts. Changing them is a governance process measured in academic years, not sprints – which is one of the reasons nobody is really at level three on their own readiness model. In the meantime, the free take-home assignment remains the backbone of assessment at most institutions, for reasons that are financial and practical before they are pedagogical: it scales, it does not need a room, and it does not need an invigilator for every twenty candidates.
The realistic path for most institutions is a combination – the take-home assignment, plus an oral defence or a sampled viva where the stakes justify it, plus similarity checking underneath all of it. In that design, the similarity report is not the verdict. It is the thing that tells a marker which conversations are worth having.
Here is the argument we would most like to see put back into the sector debate. It does not change what a similarity service is for – it still exists to find copied material – but it points to a value the discussion almost always leaves out.
A substantial share of plagiarism is not a plan. It is a first-year student who does not yet know that close paraphrase needs a citation. It is someone who reuses three pages of their own earlier assignment without referencing it, because nobody told them self-plagiarism was a category. It is a student whose prior education taught that reproducing an authority’s words is respectful rather than dishonest. It is poor note-taking that turned a quotation into what looked like their own sentence six weeks later.
This is not a hunch, and it is not special pleading from a vendor. It is the subject of our white paper series on academic misconduct, and the evidence underneath it is solid. A European survey by Mads Paludan Goddiksen and colleagues, published in Ethics & Behavior in 2023, put a set of scenarios to 1,639 undergraduates. Shown a passage of 42 words copied verbatim with no quotation marks, 37.3% failed to identify it as unacceptable. Shown the same passage with a handful of words swapped for synonyms and still no reference, that rose to 48.8% – nearly half the students surveyed. And shown genuinely good practice, properly rephrased and correctly cited, more than one in ten called it unacceptable. A Swedish study points the same way, and adds that the confusion is worst in the early years of a degree.
Read that middle figure again, because it describes the behaviour this post has been circling. The rewritten, lightly laundered passage – the thing a model now produces in seconds – is not recognised as misconduct by half the students asked. Some of the people doing it are not gaming the rule. They do not know the rule is there.
These students are not caught by assessment redesign, and they are not helped by being left undetected. They are helped by being shown, early and concretely, what academic practice requires – which is only possible if something notices in the first place.
This also disposes of the assumption buried in the post-plagiarism argument: that students who once plagiarised have simply migrated to AI, so the check is now looking in the wrong place. That assumes intent. For a large group there was never any intent to migrate, because there was never any intent to cheat. They are still making the same honest mistakes, in the same places, and they still need to be found and taught.
So the same check does two jobs. It supports the misconduct case where there is one, and it finds the students who need teaching rather than sanctioning. A similarity report read with a student in supervision, early in a programme, prevents far more misconduct than the same report read by a panel two years later. That is value an institution gives up entirely when it stops checking.
WISEflow Originality was built around this reading of the problem rather than around the AI panic.
It compares meaning, not just strings, so rewritten and paraphrased text surfaces alongside verbatim matches, in more than 50 languages. It compares students against each other – across flows, across cohorts and across institutions – because that is where the volume is, with parallel checking so that two students submitting on the same day are treated alike rather than penalising whoever pressed submit second. Matches are grouped as exact, strong or potential, so an assessor can see in seconds which reports deserve time. Institutions control their own repositories and source lists, and choose what is indexed.
And it does not guess. WISEflow Originality does not do AI detection, and we have set out why at length.
The difference is simple. A similarity match hands the assessor a document. They can open it, read it next to the submission and decide for themselves whether it is misconduct. An AI-detection score hands them a number instead, with no source to check it against – and those numbers go wrong most often for students writing in a second language. You can build a misconduct case on a document. You cannot build one on a number nobody can verify.
That is the rule we follow wherever AI touches assessment: the system flags, a person decides.
We are not post-plagiarism. What has changed is the range of things a student can do – not why they do it, not the pressures they are under, not their inexperience with academic convention, and not the fact that the easiest source of a good assignment is still another student’s good assignment.
Eaton’s actual argument is that definitions must evolve and that attribution still matters. That is a case for checking better, not for checking less.
So before removing similarity checking from your assessment stack, ask three questions. How many cases did you uphold last year, and where did the evidence come from? If the check stopped, how long before the cohort knows? And what happens to the students who are plagiarising without realising it – who finds them, and who teaches them?
Then keep the check, make it semantic, use it formatively, and put your redesign energy where it will pay off over the next five years.
Sarah Elaine Eaton, “6 Tenets of Postplagiarism: Writing in the Age of Artificial Intelligence”, 25 February 2023.
Goddiksen, M. P. et al. (2023), “Grey zones and good practice: A European survey of academic integrity among undergraduate students”, Ethics & Behavior, 34(3), pp. 199-217.
Hattingh, F. G., Buitendag, A. A. K. and Lall, M. (2020), “Systematic Literature Review to Identify and Rank the Most Common Reasons for Plagiarism”, InSITE 2020, pp. 159-182.
Karlsson, T. S. (2019), “När blir det plagiat? En tvärsnittsstudie av studenters färdigheter och förmågor gällande fusk i form av plagiat”, Högre utbildning, 9(2), pp. 32-47.