How to prove you didn't use AI on an essay
You cannot prove a negative, so stop trying to. What you can do is show how the paper came to exist, which is exactly what the procedures ask for. Swansea tells students to bring "drafts, sources and feedback". Manchester's authorship viva sets out four things it is held to confirm, and three of the four are about your process rather than your text. Build that record while you write and the question mostly answers itself.
What you are actually being asked
Nobody in an integrity process expects you to prove you did not use a tool. They expect you to demonstrate authorship, which is a different and far more achievable thing, and the published procedures say so in plain words.
The University of Manchester's humanities faculty is the clearest. Its procedure states that the viva voce for suspected malpractice "is not an academic assessment of the student's work" and that "the aim of the viva voce is to give the student the opportunity to demonstrate that the piece of work is entirely their own". It then lists what the conversation is meant to confirm, which is the whole of what you need to be able to show: that you "undertook the reading and research themselves", "undertook all the preparatory work themselves", "understands what they have written", and "wrote the piece of work themselves" (Viva Voce for Suspected Academic Malpractice).
Three institutions, three names for the meeting, one shape to the evidence.
| Institution | What it calls the meeting | What it tells you to bring |
|---|---|---|
| Manchester, humanities | Viva voce for suspected academic malpractice | "All the documentation that you used in completing the piece of work e.g. your list of source documents, evidence of primary data collection, rough notes and assessment drafts" |
| Swansea | Academic misconduct interview, and an academic integrity viva | "Evidence of preparatory work relating to the submission such as drafts, sources and feedback", presented at the interview as "any evidence which they have brought with them such as drafts, sources, etc." |
| New England, Australia | An examination on the relevant topic | Arranged "as part of the process of determination of authorship", with an optional support person who has no right of audience |
Read the right-hand column as a packing list, because that is what it is. Drafts, sources, notes, feedback. Every item on it is something that exists only if you made it while you were working, and almost none of it can be assembled afterwards in a way that survives being asked about. That is the asymmetry you are relying on, and it is a real one.
What version history proves, and where it stops
A version history is the strongest single artefact most students have, and it is also the one people most often assume they already have. Whether you do depends on where the file lived, and each of the three common answers comes with a limit its own vendor publishes.
| Where you wrote it | What is kept | The catch, in the vendor's words |
|---|---|---|
| Google Docs | A timestamped revision history of every session | "The revisions for your file may occasionally be merged." Naming a version is what stops that, and "you can add up to 40 named versions per document" |
| A .docx uploaded to Drive | Versions of the uploaded file, which is a different system | "A version might be permanently deleted after 30 days or if there are 100 newer versions", unless you mark it Keep forever |
| Word | Version history, but only in the cloud | "Version history in Microsoft 365 only works for files stored in OneDrive or SharePoint in Microsoft 365." On a personal account "you can retrieve the last 25 versions" |
A file that only ever lived on your own laptop has no history at all, which is worth knowing tonight rather than in a meeting. If that is you, the earlier drafts in your folder and their file dates are what you have, and they are still worth more than nothing.
Word's other record, Track Changes, has a limit of its own. Microsoft says Edit with Copilot "will respect Track Changes", so an AI edit is marked when tracking is on, and that Copilot "cannot turn on/off or accept/reject tracked changes", so in a file where nobody turned it on there is nothing. Tracking is a setting anyone with the file can switch off unless it is locked, which is Review, Track Changes, Lock Tracking, with a password, and a record you can turn off at will proves less than it looks like it does, because the question in the room is what happened during the stretch with no marks.
What the history shows on the other side of the desk is a shape. A guide from the University of Maryland Baltimore County tells instructors to open the version history of a submitted Google Doc and look at how the text arrived, on the reasoning that if a version "contains large blocks of text in a short period of time, one can reasonably assume the work is not original" (Using Google Doc versions to detect student originality). A document that grew over four evenings, with paragraphs rewritten and a section deleted and rebuilt, has a shape nothing else produces. That is what you are showing.
Be honest about the limit, because your marker will be. The same guide names two objections to its own method: a person who retyped pasted text would produce an ordinary-looking history, and plenty of real writing happens offline, on paper or in another editor, and never appears in the document at all. So version history is evidence of a process, not proof of authorship, and overselling it as proof is the fastest way to lose the room. Present it as what it is, and let it be corroborated by everything else. If the copy-and-paste question is the one you are worried about specifically, what Google Docs shows about pasted text goes through exactly what the history does and does not record.
Drafts, sources and the reading trail
Everything else on the packing list is cheap to keep and impossible to reconstruct. Three kinds, in the order they tend to matter.
Dated drafts. Save a copy at the end of each working session, named by date, and keep the bad ones. A tidy final file and nothing else is a weaker position than a folder of messy intermediate files, because the messy ones carry the false starts, the paragraph you cut, the argument that changed halfway. Nobody manufactures those convincingly, which is precisely why they persuade.
A source list that resolves. This is the part students underrate, and it is the part that can turn a soft allegation into a hard one. A bibliography where every entry opens is a checkable fact about your work, and a bibliography with two references that do not exist is a factual finding a marker can make in ten minutes without any judgement call about your style. That failure mode has its own mechanism, which we set out in why AI makes up citations, and you can check a finished list yourself with our free citation checker. Do it before you submit, not after somebody asks.
The reading trail. Downloaded PDFs with their file dates, annotated papers, library loan records, the tab you kept open for a week. Manchester's procedure tells students they will be asked "how you found the sources quoted in your bibliography / reference list", which is a question the trail answers and memory does not. The strongest version of this is the habit rather than the folder: record where each claim's evidence came from, with the line that supports it, at the moment you write the claim. Then your paper carries its own trail and you never have to rebuild one.
The conversation nobody can outsource
Documents get you into the room in good shape. What usually settles it is whether you can talk about your own paper. Manchester tells students how to prepare in one sentence, and it is the most useful sentence on this page: "You must be thoroughly familiar with the content of your work and the research that was involved as you will be asked questions by the members of academic staff about the sources you used, your ideas, the words used, your research, references and data collection."
So rehearse the things only the author knows. Why the argument is framed this way rather than the obvious other way. Which source changed your mind and what you believed before it. What you cut, and why it was not working. Where a particular number came from and what the paper it came from was actually measuring. Any word in your draft you would not use out loud, because "the words used" is on that list for a reason and a term you cannot define is the single most quoted thing that goes wrong in these meetings.
Two procedural facts worth having in advance. You can usually bring someone: Manchester says you "may be accompanied by a person of your choice (normally a fellow student, member of staff or Students' Union representative)", though they cannot answer on your behalf. And the meeting has a real exit. The first of its three listed outcomes is that the staff "confirm that they accept that the student wrote the work in question; no further action will be taken". People forget that the good outcome is one of the printed options.
If you are already in this and the letter arrived today, the first hour is a different job from this page, and what to do when you are accused of using AI covers it in order.
Do not lead with the detector
If a score started this, it is tempting to open by arguing about the tool. Resist it for one turn. Arguing accuracy first reads as deflection, and your process evidence is stronger ground than anyone's statistics. Make your account first, then note calmly, once, what the tool's own maker says about what the number means.
The number itself is worth understanding before you mention it, and we have that written down rather than repeated here. What a Turnitin AI score actually measures covers what the percentage counts, and will AI detectors flag my writing covers why writing that is entirely your own can raise one. If your course allowed AI with a declaration and you made one, that declaration is part of your record too, and what to write in an AI disclosure shows what a usable one looks like. Whether what you did needed declaring at all is a policy question that varies by module, which we work through in is using AI to write essays cheating.
What our own record holds
We built CiteOwl around this problem, so here is what it keeps and what it does not, stated rather than implied.
Every edit the agent makes lands as a pending change holding both the old text and the new, grouped by the run that produced it, and you accept or reject each one. Rejecting actively reverts it rather than leaving you to undo by hand. The part that matters for a source list is separate from the diff: a cited claim carries the quote from the paper it came from, so the sentence and the evidence for it sit in the same place and the check happens while you write instead of the week somebody asks. Checkpoints you star are kept for as long as the document exists.
Three costs, because a record you misunderstand is worse than none. Automatic snapshots are capped at 50 per document and the oldest are pruned past that, so a starred checkpoint is the one you can count on. Your own typing applies directly with no pending change, because only the agent proposes. And the history lives in the app, not in the file you hand in: an export carries the finished text and not the trail behind it. No tool's internal history is a document you can hand somebody, ours included, so keep your own dated copies exactly as you would anywhere else.
Build it now, not the night you need it
Everything on this page costs a few seconds a session while you are working and cannot be bought afterwards at any price. Five habits, in the order they pay off.
- Name a version at the end of every session. It is one click, and in Google Docs it is the only part of the history you are promised will survive intact.
- Write in one place, in the cloud. A document that moved between three apps has its record scattered across systems that each prune on their own schedule.
- Keep the messy drafts. Dated copies, including the ones you are embarrassed by. Those are the persuasive ones.
- Record the evidence with the claim. The source and the line that supports it, at the moment you write the sentence, not at the end.
- Keep your chat logs if you used a chatbot at all. Deleting them is the move that turns a small case into a serious one, and it is the first thing people reach for.
None of this is about beating anything. It is the ordinary shape of work that was actually done, written down as it happens, and the reason it answers the question is that it is true. A draft whose citations resolve when someone opens them, and whose edits you read before you kept them, is a paper you can talk about in any room you are asked into.