AI, pasted text, and keeping group work honest

Outsourced and pasted work is not a new problem, but it is a growing and changing one. The honest position is to be neither alarmist nor naive: take the integrity question seriously, design around it, and be clear about what a tool can and cannot do.

The problem is real, the numbers are slippery

A systematic review pooled 71 samples from 65 studies going back to 1978, covering 54,514 students. Across all of them, 3.5 per cent admitted paying someone else to do their work. In samples from 2014 onward, that figure is <strong>15.7 per cent</strong>, and the rise over time is statistically significant. ¹

Two things stop that being the headline it looks like. The first is the author’s own qualification: self-reported cheating of <em>all</em> kinds rose over the same period, the two trends correlate strongly, and the slopes are not significantly different — so contract cheating may be increasing no faster than cheating in general. ¹

The second is that the evidence underneath is weak, and the review is unusually direct about it. Seven in ten samples were convenience samples. Across those reporting enough detail to calculate one, the response rate was <strong>5.9 per cent</strong>. More than half never told students their answers were anonymous, three quarters do not report ethical approval, and fewer than a third piloted their survey. ¹

That all points one way, and it is worth following the argument rather than just noting it. People who volunteer for surveys skew older, female, better educated and wealthier — close to the opposite of the profile associated with cheating. Students who have paid for an essay have the most to fear from a survey that never promised anonymity. ¹ The likeliest reading is that <strong>15.7 per cent is a floor, not an estimate</strong>. The right tone for this whole topic follows from that: a real and probably growing problem, and not a precise statistic to brandish. The same applies to AI-generated text, which makes pasting a plausible-looking finished section trivial.

What actually reduces it

The evidence points away from suspicion and toward environment. A survey of over 14,000 students linked contract cheating to dissatisfaction with the teaching environment and a perception that there were many opportunities to cheat, and found strong student–staff relationships to be protective. ² There is no silver bullet, only better conditions. The same review’s policy recommendations put face-to-face assessment first, alongside better education and support for students and staff and changes to the law. ¹ We should be clear that a contribution record is none of those things.

Where visible authorship helps

This is where a contribution record earns its place, with limits worth stating precisely. Dwixel cannot judge whether a sentence was written by a person or a model. What it can do is show how the work came to exist: who authored which parts, and over what time.

That is worth something because of what the field currently lacks. The same review notes there are no reliable objective measures of contract cheating at all, that essay mills advertise their work as plagiarism-free and therefore invisible to similarity checkers, and that <em>objective behavioural measures are clearly desirable</em> — naming authorship-comparison tooling as a promising direction. ¹ A record of how a document was written is one such measure. It is also, on its own, weak evidence, and the rest of this section is about why.

The shape of a large paste is genuinely distinctive. In the study of collaborative authorship this rests on, edits that arrive as a single burst show up as <em>one-version insertions</em> — perturbations in the flow of a document that otherwise grows steadily. ³ A section that simply appears, fully formed, looks different from one written, and Dwixel de-credits large pasted bursts so that it does not register as authoring.

What that shape does not tell you is whether anything is wrong. The same paper offers two examples of pasted text: the Windows 98 readme dropped into a chemistry article, and long passages of scripture added to a page about Islam. ³ One was irrelevant, one was on-topic, and the insertions looked alike. Only reading the content separated them, and the authors say plainly that they could find no crisp computable definition. ³ A large paste is a reason to look, never a finding.

There is also an obvious way around it, and we would rather name it than let someone discover it. Authorship in this kind of analysis is matched at sentence level, and the authors note that a change as small as adding a comma re-attributes a whole sentence. ³ Paste a section and then edit it lightly, and much of it reads as yours. The record still shows text arriving in a burst and being touched immediately afterwards, which is itself a legible pattern, but nobody should imagine the de-crediting is difficult to defeat.

And the frozen record at the end

When a group hands in, the work is frozen into a locked, timestamped submission. We used to describe that as the foundation UK sector guidance points to for authorship verification. Having read the guidance, that was wrong, and the truth is more useful. What UK guidance actually recommends for checking authorship is the viva voce, alongside linguistic analysis tools, interrogation of document metadata, knowing your students well enough to notice change, and above all properly resourced and trained staff — which it calls the single most important step a provider can take. ⁴ A contribution record is not on that list. The same study is a useful corrective here: its most striking individual contributor was anonymous, edited one page fifty-five times over seven months, and produced substantial work that later authors kept. ³ An unusual pattern of working is not evidence of anything on its own.

Where a durable record does appear in that guidance, it is on the student’s side of the table. Setting out how an accused student should be treated, it says they <em>should be allowed to present evidence, such as date-stamped draft copies of their work, to support their claim that they did complete the work themselves.</em> ⁴ That is the honest description of what a frozen, attributed history is for. Not a way to catch someone. A way for someone accused to show their work — which is much harder to produce after the fact than an accusation is to make.

The same guidance recommends building coursework around checkpoints and early drafts, and lists iterative submission of multiple drafts among the good practice its members reported. ⁴ A record kept as the work happens is a way of doing that without adding hand-in dates.

Two cautions from the same document belong here. It is explicit that it is guidance only, not mandatory and not part of the Quality Code, so nobody should cite it as a requirement. ⁴ And more seriously: some providers report that black, Asian and minority ethnic students are overrepresented in their academic misconduct cases. ⁴ Any tool that makes suspicion easier to act on will act on that unevenly unless someone is watching for it. That is a reason to treat a contribution record as the start of a conversation rather than the end of one.

The wider sector framing is worth keeping, and the UK’s Academic Integrity Charter puts it in one line: <em>detection and penalties are important, but they cannot provide the whole solution.</em> ⁵ It sets out five things a provider should do together — educate and support staff and students, limit opportunities, deploy institution-wide detection, collect case data to improve practice, and state institutional values clearly. ⁵ Detection is one of the five, not the strategy.

The Charter is not soft on students, and neither are we. It says plainly that students are responsible for the integrity of their own learning and that the decision to use an essay mill is ultimately their own. ⁵ The argument for supporting rather than policing is not that cheating is somebody else’s fault. It is that policing alone does not work, and that a provider is better placed to give students the means to produce authentic work in the first place. ⁵ A visible, attributed record serves that goal without sliding into surveillance, and both QAA documents are clear that working in partnership with students is integral rather than optional. ^{4 5}