How to Prove You Didn't Use AI: Evidence That Actually Works

You cannot mathematically prove a negative — and the good news is that you don't have to. UK misconduct procedures work on the balance of probabilities, and published university guidance is consistent that a detector score should not be the sole evidence against you. What decides these cases is process evidence: the trail your real writing left behind, and your ability to talk about the work. This guide covers what that evidence is, in order of persuasive weight, and how to present it.

The direct answer

Assemble evidence of process, not just the finished text. A finished document says nothing about how it was made; a version history spanning two weeks, rough notes, reading lists, early drafts with mistakes that later disappear, and a confident conversation about your argument together say a great deal. Investigators are trained to look for exactly this — the Office of the Independent Adjudicator's good-practice framework requires allegations to be properly evidenced, and detector scores alone rarely meet that bar.

Detectors also genuinely misfire. Independent research found detection accuracy falls sharply on edited text, and a Stanford study reported detectors misclassifying over half of essays by non-native English speakers (Liang et al., 2023). Formulaic structure, formal register and heavy revision can all flag genuinely human writing. You are not arguing against certainty; you are adding context to a probabilistic signal.

Evidence that carries weight, ranked

1. Version history. Google Docs, Word Online and OneDrive record edits with timestamps. Dozens of sessions over days or weeks, with visible reworking, is the strongest process evidence there is — it is contemporaneous and hard to fake.
2. Drafts and notes. Outlines, mind-maps, annotated readings, earlier drafts with crossed-out ideas, supervision or feedback emails referencing the work in progress.
3. Your research trail. Library loans, saved PDFs with highlights, browser reading lists, reference-manager entries added over time.
4. Your ability to discuss the work. Being able to explain why you structured the argument as you did, define your terms, and extend a point beyond what is on the page. Many institutions treat a viva-style conversation as the decisive check.
5. Writing-style continuity. Marked earlier work in your voice — same habits, same strengths and weaknesses — gives a reviewer a baseline.
6. A dated self-scan report. Weakest of the six on its own, but useful corroboration; see below for the honest way to use one.

Present the evidence in this order. Lead with process, close with corroboration.

How to export your version history

Google Docs: File → Version history → See version history. Named and dated versions are listed with editors; screenshots of the timeline plus two or three meaningfully different versions exported as PDFs make a compact bundle.

Word / OneDrive / SharePoint: open the file online → the filename dropdown → Version History, or File → Info → Version History in the desktop app when the file lives in OneDrive. Export copies of two or three stages.

Plain files on your own machine: the file system records modification dates, and earlier saved copies (drafts folder, email attachments to yourself, backups) establish a timeline. Weaker than live version history, still worth collecting.

A dated self-scan as supporting evidence

An AI detection scan you run yourself cannot prove authorship — no detector can, in either direction, and our methodology page is explicit about that. What a dated, sentence-level report can do is show you engaged with the concern seriously: which passages carry signal, which read as ordinary human prose, and what the score was on a stated date with a stated tool. Paired with version history it corroborates your account; emailed to yourself, it is timestamped. Treat it as one document in the bundle, never the headline.

What not to do

• Do not run your work through a "humaniser" or paraphrasing tool. It is exactly the behaviour integrity procedures treat as evasion, several universities name it explicitly in their rules, and Turnitin reports paraphrased AI as AI.
• Do not fabricate drafts after the fact. Created-date metadata makes backdated files easy to spot, and a fabricated document converts a defensible case into real misconduct.
• Do not counter one score with another headline score. Detectors disagree with each other routinely; trading percentages concedes that percentages decide the question. Process evidence is your ground.

If you are already facing an allegation

Read your institution's procedure before you respond — our guide to UK academic-misconduct procedures and AI evidence walks through the typical stages, and the UK university policy tracker links what your institution has publicly said about AI detection, which is worth knowing before any meeting. Use your students' union advice service; they attend meetings and know the local process. Respond in writing, attach your evidence bundle, and stay factual.

Habits that protect your future work

Write in an environment that keeps version history switched on; keep notes and outlines even when they feel disposable; save drafts at milestones rather than overwriting one file; and if you use any permitted assistance (grammar tools, translation help), note where and disclose it if your course requires disclosure. Five minutes of habit removes weeks of anxiety later.

Frequently asked questions

Can I definitively prove I wrote something myself?

No — and neither can a detector definitively prove you didn't. Misconduct decisions are made on the balance of probabilities, which is why contemporaneous process evidence, not certainty, is what wins these cases.

Can I fail an assignment on a detector score alone?

Published UK guidance points strongly against it. The OIA requires properly evidenced allegations, and universities with public positions consistently state a score alone is not a reliable basis for a finding. If a decision cites only a score, that is a ground to challenge through the appeals process.

Does Google Docs version history count as evidence?

Yes — it is among the most persuasive evidence available, because it is timestamped, continuous and created at the time of writing rather than reconstructed afterwards.

Should I scan my own work before submitting it?

It can help, with the right expectations: a dated report showing sentence-level context documents that you checked, and gives you early warning of passages likely to draw attention. It is corroboration, not proof.

Sources and further reading

• Office of the Independent Adjudicator: Good Practice Framework
• Liang et al. (2023). GPT detectors are biased against non-native English writers. Patterns
• Weber-Wulff et al. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity
• UCL: GenAI and academic integrity
• Genutext: AI detector false positives · UK misconduct procedures · UK university policy tracker · Accuracy methodology