How to Prove You Wrote It Yourself (Before Anyone Accuses You)

The short answer

You prove you wrote something by showing the process, not by arguing about the text. Keep your document's version history intact, write in one place rather than pasting from many, keep dated notes and annotated sources, and save drafts at natural milestones. A detector score is a claim about the finished words and cannot be argued away. A version history showing weeks of incremental writing, false starts and revisions is evidence a detector cannot produce and an accusation struggles to survive.

The uncomfortable part: this evidence cannot be created afterwards. It is only worth something if it already exists when the accusation arrives — and students who get accused almost never expected to be.

Why this is worth doing before anything happens

If you spend any time in student forums you will notice how often the same post appears: a paper someone wrote entirely themselves came back flagged, and now they are trying to prove a negative to a committee. The most useful advice in those threads consistently comes from other students, and it is always some version of the same thing — start saving your drafts.

That advice is correct, and it is worth understanding why it works:

• You cannot disprove a score. A percentage is not an argument you can meet with a counter-argument. There is nothing to rebut.
• You can demonstrate a process. Writing leaves a trail — dead ends, restructured paragraphs, a sentence rewritten six times. Generated text does not have that history, and the absence is noticeable.
• It shifts the conversation. An accusation based only on a detector score asks you to prove innocence. Process evidence turns it into a review of documented facts.

It costs almost nothing. If you already write in Google Docs or Word, most of it happens automatically and your only job is to not destroy it.

What to keep, exactly

In rough order of how much weight each carries:

1. Version history. Google Docs keeps this indefinitely; Word keeps it when the file lives in OneDrive or SharePoint. This is the strongest single artefact you can have, because it is timestamped, automatic and granular.
2. Dated research notes. Notes taken while reading, with dates, showing ideas forming before they appear in the text.
3. Annotated sources. Highlighted PDFs and margin notes. These are difficult to fabricate convincingly and show genuine engagement.
4. Milestone drafts. Save a dated copy after each chapter or major revision —thesis-v1-2026-03-14.docx and so on.
5. Supervisor correspondence. Emails with feedback on earlier versions are third-party confirmation that the work existed in a rougher state.
6. A short process note. A paragraph recording which tools you used and for what — spellcheck, reference manager, translation, anything. Written before any accusation, it reads as candour rather than defence.

How to set it up in five minutes

• Write in one document from the start. The most common way students destroy their own evidence is drafting in five places and pasting a finished text into a new file. That file has no history and looks exactly like generated text would.
• Do not paste large blocks. Where you must — moving your own text between documents — keep the source file.
• Check that history is actually on. In Google Docs: File → Version history. In Word: the file must be in OneDrive or SharePoint, not local-only.
• Keep one dated backup per milestone somewhere separate from the live file.
• Do not delete anything until your degree is confirmed, including the messy early drafts. Especially those — they are the most convincing part.

If you have already been accused

The evidence you have is the evidence you have; do not manufacture more, and do not edit anything now. Retroactively created material is usually detectable and turns a survivable situation into a serious one.

Work through this in order:

1. Ask what the accusation actually rests on, in writing. If it is only a detector score, that matters — most vendors state their score is not sufficient for a finding on its own, and their documentation is usually your strongest citation.
2. Export your version history before doing anything else to the document.
3. Assemble a timeline: notes, drafts, emails, source annotations, in date order, with the final text at the end.
4. Ask for the false-positive rate of the tool used. If nobody can supply one, that is itself a substantive point about the strength of the evidence against you.
5. Find your institution's appeals procedure and its deadline. Deadlines here are short and are enforced.

We have a longer walkthrough of that conversation in what to do when you are accused of using AI, and on what a flag actually means in an AI detector flagged my thesis.

What does not work as evidence

• “Run it through a different detector.” Detectors disagree constantly. A second opinion that clears you is easy to dismiss, and one that also flags you makes things worse.
• Rewriting to sound less like AI. This damages your writing, and if discovered mid-process it looks like concealment. It is also the opposite of what a good writing centre would tell you.
• Insisting you would never do that. True and unpersuasive. Committees hear it from everyone.
• A screenshot of a word processor. Shows nothing about how the text got there.

What detector percentages actually mean

A detector score is not the probability that AI wrote your text, and it is not the proportion of your text that is AI. It is a measure of how closely your writing resembles patterns the tool associates with generated text. Formal, structured, hedged academic prose resembles those patterns — which is why the writing style universities explicitly teach is the style most likely to be flagged.

That is also why the number that matters is not the score but the false-positive rate: how often the same verdict lands on writing known to be human. Almost no vendor publishes one. We publish ours next to every verdict, and the measurements and corpora are on our accuracy page — including the cases where we refuse to give a verdict at all because we could not make one honest.

If you want to know what a detector sees in your own writing before your university runs one, that is what our free check is for. It will not tell you that you are safe — no tool honestly can — but it will show you which passages a detector reacts to, and how often that reaction is wrong.