Last updated: 2026-10-18
Why generative AI changes the form and the conditions of academic misconduct without creating the problem, and what that means for assessment
Revised following an evidence-trace review (17 October 2026). The claims on this page were tested against their sources using the method described in Thinking in Public. The case study Has Generative AI Changed the Problem of Cheating, or Only Its Form? records what held, what did not and what remains open. Passages marked Revised were changed as a result, and the earlier wording and the reasoning for each change can be opened beside the passage. The changes: the prevalence figures are re-sourced and qualified; the thesis now says what generative AI has changed; the evidence on detection is described precisely and extended to its fairness problem; the account of why students cheat adds unmet need and corrects a description of the staff survey; and the claims about mixing assessment methods and about suspicion are qualified.
Earlier wording
Why generative AI changes the form of academic misconduct without creating the problem, and what that means for assessment
Why it changed
The thesis now says that generative AI changes the conditions of misconduct as well as its form. See evidence trace 7.
Generative AI has made assessment impossible, the argument goes. A student can ask a chatbot to write the essay, the report or the code, and no coursework is safe. The only answer, on this view, is a return to examinations, vivas and tightly controlled conditions.
The difficulty with that argument is not that AI creates no new opportunities for misconduct. It clearly does. The difficulty is the assumption underneath it: that before AI, students could not cheat, and that assessment was a secure system which a new tool has broken. That assumption does not survive contact with the history of higher education, and it leads to a poorer set of questions than the ones we should be asking.the history of cheating is long and varied
This page takes a different position. The challenge is real, but it is not unprecedented. Cheating has always mattered, and educational systems have always had to balance trust, verification, learning, assessment and ethics. Generative AI changes the tools available to a student. It does not remove that balancing act, and it should not be allowed to hide it, but it does change two of the things the balance depends on, as the section on what AI changes explains. Revised
Earlier wording
Generative AI changes the tools available to a student. It does not change that balancing act, and it should not be allowed to hide it.
Why it changed
The balancing act remains, but two things it depends on have changed: outside help in unsupervised work is now cheap and rarely caught, and the line between misconduct and ordinary use has blurred. See evidence trace 4 and evidence trace 7.
The forms of misconduct are familiar to anyone who has taught or studied for long. In examinations, students have copied from neighbours, hidden notes, used materials they were not allowed to bring, and received help from people outside the room. In coursework, students have plagiarised, colluded, bought finished work, and resubmitted essays written for another module. In projects, students have exaggerated their own contributions, borrowed code, and presented the effort of a group as the effort of one person.
Revised Survey evidence from American universities gives some sense of the scale. In 1963, Bowers surveyed students at 99 institutions, and McCabe and Trevino repeated the survey in 1993 at nine of those schools. At those schools the proportion of students who admitted copying from another student on a test or exam rose from 26% in 1963 to 52% in 1993, and unpermitted collaboration on assignments rose from 11% to 49%[1]. The overall picture is less dramatic than those two figures suggest. The proportion admitting any serious cheating rose from 75% to 82%, serious cheating on written work was almost unchanged (65% and 66%), and admitted plagiarism fell from 30% to 26%[1]. The authors describe a modest increase in overall cheating, with significant increases in the most explicit forms of test cheating and in collaboration, and suggest that students' changing view of what counts as plagiarism may explain the written-work results[1]. All of these are self-reports from different samples (452 students in 1963 and 1,793 in 1993), not a controlled trend, and they describe a period before generative AI existed. They show that misconduct was common long before a chatbot could write an essay, and that some forms of it had grown.
Earlier wording
Survey evidence from American universities gives some sense of the scale. McCabe and Trevino's study of undergraduates reported that the proportion who admitted copying from another student on a test rose from 26% in 1963 to 52% in 1993. Unauthorised collaboration on work meant to be done individually rose from 11% to 49% over the same period (McCabe & Trevino, 1993). Those figures are about self-reported behaviour, and they are about a period before generative AI existed. They show that misconduct was common, and that it was growing, long before a chatbot could write an essay.
Why it changed
The figures were accurate, but they come from McCabe, Trevino and Butterfield (2001), which compares Bowers's 1963 data with a 1993 replication at nine of his schools, and not from the 1993 paper cited, which reports a different survey. The same tables show a modest overall rise, almost unchanged cheating on written work and falling plagiarism, so "it was growing" holds for explicit test cheating and collaboration and is too strong as a general statement. See evidence trace 1.
Revised Research across 31 campuses in 1990 and 1991 found that students' perceptions of how faculty and other students behave were the most significant factor in how much cheating they reported[2][3]. That is an association rather than proof of cause, and it is a point about culture as much as opportunity. It returns later in this page.
Earlier wording
The same study found that students were more likely to report dishonest behaviour when they believed their peers were doing it too (McCabe & Trevino, 1993). That is a point about culture as much as opportunity, and it returns later in this page.
Why it changed
The finding comes from the 1990-91 multi-campus survey, and the later summary describes peer culture as the most significant factor. It is an association; the sources do not show which way the cause runs. See evidence trace 6.
The conclusion is simple. Assessment has never operated in a world where misconduct was impossible. Every system that rewards a credential has also rewarded the people who found ways around it.
The usual proposal in response to AI is to return to invigilated examinations. That proposal has real strengths, and it deserves a fair hearing before its limits are discussed.
Examinations verify identity and limit outside assistance during the test. They also produce a result under known conditions, which makes comparison between students more straightforward. Their weaknesses are equally real. They can produce stress that has little to do with the capability being assessed. They test performance under artificial constraints, which may not match how the work is done in practice. They can disadvantage students with particular access needs, and they reward the skills of rapid recall, which matter less than they once did.
Coursework allows authentic tasks, time for reflection, and richer outputs than a timed paper can produce. It also creates real opportunities for misconduct, including outsourcing, collusion and reuse of earlier work. The trade-off between authenticity and verifiability is built into the method.the spectrum of academic misconduct
Projects reflect professional practice and integrate skills across a programme. They are also difficult to attribute. When a team produces the work, it is hard to say with confidence what each member did, and that difficulty is a form of opportunity in its own right.
The key message is that no assessment method completely eliminates misconduct. Every method involves trade-offs between authenticity, verification, cost and fairness. Choosing an assessment is choosing which problems to accept, and the choice should be made with that in view.
It helps to look at how other systems handle misuse. Financial systems have fraud. Online platforms have abuse. Organisations face insider threats from staff who have access they should not misuse. In none of these cases do we expect perfect prevention, and in none of them is the absence of perfect prevention treated as proof that the system is broken.
What organisations do instead is reduce opportunity where it is cheap to do so, detect problems in proportion to their seriousness, build cultures in which misuse is less likely and easier to report, and manage the residual risk. The aim is proportionate assurance, not an impossible guarantee.
Education works in the same way. The goal is not a system in which cheating is impossible. The goal is a system that gives reasonable confidence that a qualification means what it says, at a cost that the institution and the students can sustain. A standard of perfect security would be expensive and would damage the relationship that makes most learning possible.cf. the 'productive struggle' page
A page about misconduct that only discusses procedure misses the point. People cheat for reasons, and most of those reasons predate any tool.
Revised Unmet need belongs on the list too. A survey of 14,086 students at eight Australian universities found three factors significantly associated with contract cheating: dissatisfaction with the teaching and learning environment, a belief that there are many chances to cheat, and speaking a language other than English at home[4]. The last is not evidence that these students are less honest. It points to unmet needs for language and learning support as one route to outsourcing.
The survey evidence linked contract cheating to language background and unmet support needs, a factor the original list left out. The sentence says what the finding does not show. See evidence trace 5.
Survey research on academic staff suggests that the main obstacle to acting on misconduct is proving it. Keith-Spiegel and colleagues asked a national sample of psychology instructors why they overlooked evidence of cheating. The most frequent reason was insufficient evidence that cheating had occurred. Their analysis found four other groups of reasons: the time and effort needed to deal with it, fear of retaliation or a legal challenge, emotional reasons such as stress and lack of courage, and denial, such as the belief that cheaters would fail anyway[5]. Revised Whitley and Keith-Spiegel's guide to the subject treats academic dishonesty as a problem for the whole institution, not only for the student who commits it[6].
Earlier wording
Keith-Spiegel and colleagues asked a national sample of psychology instructors why they overlooked evidence of cheating. The most frequent reason was insufficient evidence that cheating had occurred, followed by the time and effort that dealing with it required (Keith-Spiegel et al., 1998).
Why it changed
The abstract of the study names insufficient evidence as the most frequent reason, and then four other groups of reasons without ranking them. "Followed by the time and effort" was not supported. See evidence trace 5.
The conclusion I draw is that AI did not create these motivations. It provides another mechanism through which they can be expressed, and it makes some of them easier to act on. That is a real change, and it does not change the motivations themselves.motivation is the why; AI is just the how
This is the point where the page connects most strongly to the rest of this section. A student who sees education primarily as mark acquisition will naturally look for the most efficient route to the mark. A student who sees education as capability development has far less reason to outsource the learning, because the outsourced work would not build the thing they came to build.capability is the muscle; marks are just the ruler
Revised The research supports an association here. A meta-analysis of 33 studies found that a focus on mastery and competence is negatively associated with academic dishonesty, while the findings for performance goals are mixed. Most of those studies were correlational and relied on self-reports, so the direction of cause is not established[7].
The claim that a learning focus reduces cheating is supported as an association, and the studies behind it are mostly correlational and self-reported. See evidence trace 5.
That difference is partly about incentives and partly about identity. Intrinsic motivation, a sense of professional identity, and an understanding of what the qualification is for all reduce the appeal of shortcuts. They are not produced by warnings. They are built by assessment that rewards the capability, by teaching that makes the purpose visible, and by a relationship in which a student's effort is noticed. The earlier discussion of trust and the staff–student relationship makes the same point from a different direction.
The real victim of cheating is rarely the institution. It is the student who avoids developing the capability they will later need, and who finds that out at the worst possible moment, in a job that expects them to know what their qualification says they know.
Universities tend to respond to misconduct by policing behaviour, detecting it and punishing it. Those responses have their place. They are not sufficient, because they only address the behaviour that is visible to the institution.
A complementary aim is to help students develop honesty, responsibility, professional judgement and accountability. A graduate who behaves ethically only when they are being monitored has not yet become an ethical professional. They have learned what the monitoring looks like. The difference matters most in the situations no monitor will ever see, which is where most professional work happens. Whitley and Keith-Spiegel's treatment of integrity as a system-wide question, rather than a matter for individual enforcement, points in the same direction[6].
Only now do I want to move explicitly to generative AI, because the historical point has to come first. There are genuine changes, and the concerns about them are legitimate.
Earlier wording
Traditional similarity checks compare text against existing sources, and a newly generated text may match nothing. The evidence on detection is discouraging. In a blind test at a UK university, entirely AI-written submissions went largely unflagged by the standard processes (Scarfe et al., 2024).
Why it changed
The blind test added fully AI-written scripts, about 5% of the marking in each of five modules, to real marking, and 94% went undetected. The earlier wording left out the design. Two further results bear on detection: dedicated software is not accurate or reliable, and its errors fall most heavily on non-native English writers. See evidence trace 4.
These are changes in degree, and in some respects in kind. They are not, however, changes in the underlying educational problem. That problem is the one this page started with. How do we create assessment systems that support learning while keeping reasonable confidence in what students have achieved? That question was being asked before generative AI, and the earlier answers to it, such as the contract-cheating research[4], were already pointing at the same causes: teaching environments, assessment design and opportunity.
Revised Two of the inputs to that problem have nonetheless changed. Outside help in unsupervised work used to cost money and carry some chance of being caught. It now costs little or nothing and, as above, is rarely caught. And the boundary between misconduct and legitimate use has blurred. In a UK survey of full-time undergraduates, 88% said they had used generative AI for assessed work in the 2025 survey[11], and 94% in the 2026 survey[12]. When almost every student uses these tools, part of the problem is no longer misconduct. It is whether a grade still shows the capability it claims to show, a point made in the argument that validity matters more than cheating[13].
The page argued that the underlying problem was unchanged. The evidence supports that for the balancing act, but outside help is now cheap and rarely caught, and most students now use AI for assessed work, so part of the problem has become one of validity. Both HEPI surveys are cited. See evidence trace 2, evidence trace 4 and evidence trace 7.
There is no single solution, and I would be suspicious of anyone who offered one. A better approach is to combine several forms of evidence, so that no single assessment carries all the weight. Coursework, projects, reflective commentary, demonstrations, discussions, presentations, practical work and, where appropriate, examinations can all contribute.
The advantage of this diversity is that each form of evidence is weak in a different way, and the weaknesses do not line up. A student who has outsourced a written report may find it much harder to answer a question about it in a viva. A student who can explain a design in conversation may still need practice to produce it under time pressure. Several kinds of evidence together give a richer picture of learning than any one of them alone, and they plausibly make misconduct harder to sustain across the whole programme, although I have found no study that tests that directly. The design principles for this are set out in Teaching with GenAI, and the question of what assessment should reward is discussed in Generative AI in the Curriculum.
Earlier wording
Several kinds of evidence together give a richer picture of learning than any one of them alone, and they make misconduct harder to sustain across the whole programme.
Why it changed
No study found tests whether mixing assessment forms makes misconduct harder to sustain, so the claim is now stated as plausible. See evidence trace 3.
Revised Two cautions apply. Authentic tasks are not a guarantee: an analysis of 221 orders placed with essay-writing sites and 198 tasks in which contract cheating was detected found that tasks with none, some or all of five authenticity features were routinely outsourced[14]. And rules about permitted AI use are not enough on their own. It has been argued that most responses rely on instructions students can ignore, creating an "enforcement illusion", and that the more dependable lever is structural change to what the task requires[15].
Authentic tasks were routinely outsourced in an analysis of detected contract cheating, and rules about permitted use change little unless the task itself requires the capability. See evidence trace 3.
An educational system built entirely on distrust would be difficult to sustain. Students need clear expectations and clear accountability, and they also need to be trusted as developing professionals who can be expected to do their own work. The research on campus culture points the same way: peer culture was the most significant factor in how much cheating students reported, and strong cultures of integrity were found on campuses with and without honour codes[3]. Whether suspicion itself produces a culture of misconduct has not been tested directly. A system built on suspicion is nonetheless unlikely to build the culture that does the work, and it risks making the honest majority less willing to report the dishonest minority. Revised
Earlier wording
A system that treats every student as a suspect will produce the very cultures in which misconduct is normal, and it will make the honest majority less willing to report the dishonest minority.
Why it changed
Peer culture predicts misconduct, but no source consulted tests whether suspicion produces a culture of misconduct, so the causal claim is now qualified. See evidence trace 6.
The question is therefore not whether trust should exist. It is how trust and verification should be balanced. That balance is a judgement about proportion: how much evidence is needed for each decision, what the consequences of error are, and who bears them. The earlier page on Supporting Students with GenAI argued that relationships are part of that balance, and that they cannot carry the whole of it.
Students have always been able to cheat. They could cheat before AI, before the internet, before coursework was common, and in examinations as long as examinations have existed. Generative AI does not create the problem of academic misconduct. It changes its form, sometimes dramatically, and with it two of the conditions the problem depends on: help is now cheap and rarely caught, and the line between misconduct and ordinary use has blurred. It deserves a serious response. Revised
Earlier wording
Generative AI does not create the problem of academic misconduct. It changes its form, sometimes dramatically, and it deserves a serious response.
Why it changed
The conclusion now includes the two conditions that have changed, in line with the revised thesis. See evidence trace 7.
The challenge for educators is therefore not to design assessments that are impossible to cheat. Such assessments have never existed. The challenge is to create learning environments that develop capability, encourage integrity, reward genuine understanding, and provide reasonable confidence that students have achieved what their qualifications claim.
That was the challenge before generative AI. It remains the challenge today. The useful shift in the conversation is from asking how we stop students using AI to asking how educational systems have always balanced trust, learning, incentive and assurance, and what that balance requires now.
University of Reading Graduate Attributes framework.