Why process beats text
A detector looks at a finished document and estimates how predictable it is. It has no access to how the document came to exist, which is precisely why it cannot distinguish a careful non-native writer from a language model. The same limitation cuts the other way: nothing in a finished file can prove you wrote it either — not your name on it, not your style, not a low detector score.
Process evidence escapes that trap. A history of a document forming over days, with false starts, reordered paragraphs and abandoned sentences, describes something a generated file does not have. It is also very difficult to manufacture after the fact, which is exactly what gives it weight.
Google Docs: the strongest artefact most students have
File → Version history → See version history, or Ctrl/Cmd + Alt + Shift + H. Docs records edits close to continuously rather than as daily snapshots, so you can scrub through the document assembling itself.
Three things to know before you rely on it:
- Browsing history requires edit permission. Google’s documentation is explicit: “To browse earlier versions of a file, you need permission to edit that file.” A viewer or commenter cannot see it. So keep edit access on your own copy, and if you submitted by sharing, do not downgrade your own permission.
- Name the versions that matter. In the history panel you can name a revision. Naming two or three — “first outline”, “after library session” — makes the timeline legible to someone scrolling it for the first time in a meeting.
- Do not start a clean copy. “Make a copy” produces a new file whose history begins at the copy. Tidying up in this way is the single most common way students destroy their own best evidence.
Microsoft Word
Version history exists where the file lives in OneDrive or SharePoint and AutoSave was on: File → Info → Version History. A document that only ever existed on a local disk generally has none, though the file’s created and modified timestamps still say something, and previous versions may exist in your backup system.
If you have no version history at all
This is common and not fatal. No single item below is as strong as a full revision history, but several of them together describe a process, and a process is what you are demonstrating.
- Drafts emailed to yourself, a friend, or a writing centre — the timestamps are the point.
- Files in cloud sync or backup with modification dates spread across days.
- Photographs of handwritten notes, plans or mind maps.
- Browser history for the sources you read, and library loan or database access records.
- Messages where you discussed the assignment, complained about it, or asked a question.
- Feedback on an earlier draft from a tutor, a friend, or a writing centre appointment.
What this evidence does and does not establish
Be straight about the limits, because overclaiming in a hearing damages you. Version history does not literally prove that a human composed every word — text can be pasted in and then edited, and a determined person could stage a history. What it does establish is a pattern of gradual composition consistent with authorship and difficult to reconcile with wholesale generation.
That is a materially stronger claim than anything on the other side of the table. A detector score is a probability estimate produced by a tool that, in the largest peer-reviewed test of fourteen such tools, was found “neither accurate nor reliable”. Your evidence is a record; theirs is an inference.
One caveat we would rather give you than have you discover. Evidence only helps where someone is obliged to weigh it. Gizmodo documented a freelance copywriter who was flagged by a detector, produced timestamped documentation showing he had written the article by hand, and lost the work anyway — roughly 90% of his income. A commercial platform can drop a contractor without a hearing. An academic institution generally cannot: you have a process, a right of reply, and usually an advocate. That is why the same evidence is worth far more inside a formal procedure than outside one, and why the first thing to establish is what procedure you are actually in.
Presenting it
- Build a short timeline — dates, what existed at each point, one line each. Do not hand over a raw folder and expect it to be read.
- Offer to walk through the history live. Scrubbing a version slider in front of someone is more persuasive than a description of it.
- Be ready to talk about the argument. Someone who wrote an essay can explain why the third section comes third, what they cut, and which source changed their mind. This is the oldest and still the best test.
- Keep the evidence intact. Do not edit, re-export, or reorganise anything while a process is open.
Going forward
The cheapest insurance is simply to draft in something that keeps history, and to stop deleting your notes. That is close to the whole of it. We would not suggest anyone reorganise how they study around an accusation they have not received — but if you are in a course where detector scores are being treated as evidence, working in a document with a full revision history costs nothing and settles the question in advance.
Who you present it to
Draft history works the same everywhere; who reviews it does not. In England and Wales an independent body will look at the case after your university finishes with it — see the UK appeal process, and who regulates detection there. In India, the panel deciding your case is defined by regulation, and so is the percentage it is reading — the UGC rules explained.
Related reading
- Falsely accused of using AI — what to do — the wider process, what to say, and the institutions that stopped using these tools.
- Why AI detectors flag human writing — the mechanism, and who it hits hardest.
- Appealing an AI detection finding — what to do once a written decision exists, and how the evidence above is used at each stage.
- Sentence-level detection — why asking which passages produced a score turns a verdict into something you can actually examine and answer.
- Stylometry — the name for what you are doing when you show that disputed work matches your own earlier writing, and why a portfolio of drafts outweighs a detection score.