What Academic Integrity Means When Machines Can Write
Every school has an academic integrity policy, and most were written for a world in which the hardest way to get an essay was to write it. The policies list offences: copying, collusion, buying work, bringing notes into an exam. Underneath the list sits an assumption the policies never state, because it never needed stating: that a competent piece of student writing was, almost always, produced by the student's own thinking. A teacher who asks what is academic integrity now, with a language model on every device, is asking whether that assumption can be repaired or must be replaced.
This essay argues that it must be replaced, and that the replacement is older than the rule book. The International Center for Academic Integrity defines the concept through six fundamental values rather than a list of offences, and the word integrity itself means wholeness. Read that way, academic integrity was always a claim about the relationship between a piece of work and the person who learned by making it. Generated text changes only what counts as evidence for that claim.
The short answer. Academic integrity is the condition in which a student's submitted work and the student's own learning are one thing, honestly represented. Since machines can now produce the work without the learning, integrity can no longer be inferred from the product. It has to be evidenced through visible process and honest disclosure of any assistance received.
What is academic integrity? Six values and an older word
The most widely used academic integrity definition comes from the International Center for Academic Integrity's fundamental values, first published in 1999 and revised in 2021: honesty, trust, fairness, respect, responsibility and courage. The list reads better as a description of a community than of a student. Honesty is the student's, but trust is the teacher's to extend, fairness is the institution's to guarantee, and courage is what everyone needs when the honest course is the costly one. Academic honesty is one value among six; the other five describe the conditions under which it is possible.
The word is older than any of this. Integrity shares a root with integer: something whole, undivided, entire. Applied to learning, a piece of student work has integrity when the artefact and the understanding behind it are a single thing, when the essay is the visible surface of thinking that actually took place in the person who signed it. Plagiarism breaks that wholeness by inserting someone else's thinking; contract cheating by inserting someone else's labour. Generated text breaks it in a new way: the artefact is whole, fluent and original in the copyright sense, and yet nothing behind it belongs to the student at all.
The proxy that generated text broke
For most of the history of mass schooling, a teacher never had to check for wholeness directly, because the product implied the process. A competent essay required reading, planning, drafting and revising, with no cheap route around those steps, so integrity policies could be written as lists of the few expensive routes that existed. This publication's earlier essay on assessment when product severs from process sets out how completely that inference has failed: a fluent product is now available in seconds to a student who has understood none of it.
This is an evidentiary failure rather than a discipline problem. Offence-list rules assume that a violation leaves a trace in the product: a matching passage, a change of voice, a source that cannot be found. Generated text leaves none of these reliably. The detection tools sold to fill the gap read statistical fluency rather than authorship, and read it badly for some students: Liang and colleagues (2023) found that widely used detectors flagged most essays by non-native English writers as machine written while passing native speakers' work. A policy that keeps the offence-list form and adds a detector has bolted an unreliable witness onto a broken proxy. The companion piece on how accurate AI detectors are reviews that evidence in full.
Four kinds of academic misconduct, and what each claims falsely
Academic misconduct is a family of acts that share one feature: the student presents a false claim about the relationship between the work and themselves. What differs is the claim, and therefore the evidence that settles it. Treating undisclosed generation as a species of plagiarism, as many policies now do, sends investigators after the wrong evidence.
| Category | What the student falsely claims | What evidence settles it |
|---|---|---|
| Plagiarism | "These words and ideas are mine" when they belong to an identifiable source | Text matching against the source; the plagiarism service most universities license does this well |
| Contract cheating | "I did this work" when a paid or unpaid third person did it | Payment records, communications, and above all the student's inability to discuss the work in an oral defence |
| Unauthorised assistance | "I did this under the stated conditions" when a tutor, parent, peer or tool helped beyond what the task allowed | The task's published rules, the draft trail, and the student's account of who did what |
| Undisclosed generation | "This is the product of my learning" when a language model produced it and the student did not say so | Version history, drafts and notes, an oral defence, and comparison with the student's supervised writing; detector scores alone settle nothing |
Two things follow. The evidence that settles the newest case is the oldest evidence in education: the trail of the work and the student's command of it. And the only category a text-matching tool settles is the one it was built for. Bretag and colleagues (2019), surveying more than 14,000 Australian university students, found that roughly 6 per cent reported at least one outsourcing behaviour, associated with dissatisfaction with the learning environment. The essay on contract cheating after generated text traces how the essay mills' business has changed; the evidentiary lesson has not.
Integrity redefined: evidenced process and honest disclosure
If wholeness can no longer be inferred from the artefact, then integrity in the age of AI has to be defined by the two things that can still show it. The first is evidenced process: drafts, notes, version history, in-class writing, and the short oral defence in which a student explains the choices in their own work. Together they make it reasonable to believe that the learning happened in the person claiming it. The guide to the process portfolio as evidence of learning sets out what a workable trail looks like across subjects.
The second is honest disclosure. Once a language model is a normal tool, the integrity question is no longer whether it was used but whether its use was permitted and declared. A student who writes "I used a language model to generate three counterarguments, rejected two and rebuilt the third" has made a claim that can be checked. A student who says nothing has made a different claim, and it is the silence rather than the tool that breaks integrity. A written disclosure statement for coursework is the mechanism by which the honesty value is exercised, and it converts most cases from investigations into conversations.
Defined this way, the six values reorganise around authorship. Honesty becomes disclosure, trust becomes the teacher's willingness to accept a documented process without a detector, and fairness becomes a promise that no student will be judged on a probability score.
The apprenticeship tradition already held the answer
None of this is new. The oldest teaching relationship never had an integrity problem in the modern sense, because in a workshop the master watched the work being made. Collins, Brown and Holum's 1991 essay on cognitive apprenticeship and making thinking visible, the flagship of this site's archive, argued that school had hidden the thinking that apprenticeship kept in view, and that good instruction consists largely of bringing it back into sight. The Initiative's own paper on how apprenticeship transmits judgment makes the same point from the other side: the apprentice's competence was certified by the master's long observation of how the work came to be.
When thinking is visible, authorship is evident. A teacher who has seen a student's outline on Tuesday and their muddled first paragraph on Thursday, and then the reworked argument the following week, does not need a machine to tell her whether the final essay is the student's. Assessment that makes thinking visible was recommended for its learning effects long before machines could write; it now has a second justification pointing at the same practices.
What this means for a school's integrity policy
A policy fit for the present states the definition before the offences. It says that academic integrity means the work and the learner are whole, that the institution's job is to design assessment in which that wholeness can be seen, and that students are expected to make their process visible and their assistance explicit. It distinguishes the four categories above and names the evidence appropriate to each, and it says plainly that a detector score is a reason to ask, never a reason to conclude. It applies the same honesty to the mass of unsupervised work a school assigns, which is why the homework question, asked honestly belongs in the same conversation. A working template for an AI integrity policy follows this shape.
Frequently asked questions
What is the simplest definition of academic integrity?
Academic integrity is the honest representation of the relationship between a piece of work and the person who submits it. The International Center for Academic Integrity expresses it through six values: honesty, trust, fairness, respect, responsibility and courage. In practice it means the work you hand in reflects learning that actually happened in you, and any help you had is declared.
Is using AI a breach of academic integrity?
Not by itself. Using a language model breaches integrity when the task prohibited it, or when the use was permitted but not disclosed. A student who declares what was asked, what came back and what they kept has been honest; a student who presents generated text as the product of their own learning has not. The task's rules and the disclosure decide it.
What is the difference between plagiarism and using AI?
Plagiarism takes identifiable words or ideas from a source and presents them as the student's own, and a text-matching tool can usually find the source. Generated text has no source to match, so the false claim is different: the student presents someone else's process as their own. That case is settled by drafts, version history and an oral defence, which matching software cannot supply.
Can a school prove a student used AI?
Rarely from the text alone, and detector scores are too unreliable to carry a finding. What a school can establish is whether the student can account for the work: explain the choices in it, show the drafts that led to it, and produce supervised writing of comparable quality. Where those are absent the school has grounds; where they are present, it does not.
Where this leaves a school
The question what is academic integrity has a stable answer, and generated text has clarified rather than unsettled it. Integrity is wholeness between the learner and the work, honestly represented, and a school upholds it by designing assessment in which the process of learning is visible and by asking students to declare the help they used. Rule lists and detection software both tried to read integrity off the product, and the product can no longer be read that way. The Academic Integrity hub gathers the working guidance that follows, and the essay on authentic assessment when the product can be generated turns this definition into task design.
Two further essays in this pillar sharpen the definition at its edges: whether a teacher can detect AI-written work by reading it, and what Content Credentials add for media work where a written trail does not exist.