How to Prove You Didn't Use AI

By Jack Stovell · 2026-09-04 · Guides

You can't prove a negative. That's the annoying truth sitting under this whole problem, and no amount of righteous indignation changes it. Someone's accused you of using AI to write something you wrote yourself, and now you're stuck trying to demonstrate the absence of a thing. Logic doesn't work that way. But persuasion does, and that's the distinction that actually gets you out of this.

Here's the thing: nobody's ever demanded you prove you didn't cheat on an exam by showing an empty hand. They ask for your working. Same principle applies here, just with different tools.

Process evidence beats protestation

The most convincing thing you can hand someone is the mess. Not the polished final draft, the actual trail: version history in Google Docs, timestamped saves, the three abandoned openings you deleted at 11pm, the notes where you argued with yourself about which example to use. Real writing looks like wreckage before it looks like prose. AI output doesn't have that wreckage because it doesn't need multiple attempts to get somewhere; it just goes there.

So show your revision history. If you wrote in Word, check whether track changes or version history is still live. If you used Docs, the entire timeline is sitting there waiting, every edit stamped with a time and (usually) your account. Screenshots of a blank page slowly filling up are boring to look at and devastating as evidence. Boring is what you want.

Keep your notes too, even the scrappy ones. A bullet list of half-formed ideas, a voice memo transcript, an annotated PDF of a source you argued with in the margins. These things don't get generated after the fact for show. They're proof of the argument you had with yourself before you had it with anyone else.

Consistency evidence: your voice has a fingerprint

Here's something people don't think about until they need it: you've been writing in a recognisable way for years, and that consistency is itself a form of evidence. Not vibes, not "it sounds like me", but something closer to measurable. Sentence length, how often you use contractions, whether you reach for semicolons or avoid them, your rate of hedging versus stating things flatly. These patterns hold steady across huge stretches of your writing, the same way a signature does.

That's why pulling up old essays, old emails, old forum posts next to the disputed piece is so effective. If your last ten pieces of writing all show the same rhythm and this one matches, that's not coincidence, that's you. This is the exact idea behind stylometric analysis, and it's been used for decades to settle authorship disputes long before anyone worried about AI. It's not new science pressed into service. It's old science finally getting useful again.

If you want to see this in cold, specific terms rather than take my word for it, tools exist that measure this properly. ScriptGrain, which I build, looks at 45 separate attributes of how you write and builds a profile from it, free, no card needed for the first pass. There's also a free browser tool at scriptgrain.com/tools/writing-style-analysis if you just want a quick read on your own patterns before any dispute even happens. Knowing your own fingerprint in advance is a different kind of leverage than scrambling for it under pressure.

Live evidence: walk them through it

Sometimes the most convincing thing isn't a document at all. It's you, in the room, talking through the piece. Explain why you chose that opening line. Explain what you were trying to do in paragraph four and why it didn't quite land. Edit something live, in front of them, and let them watch you make a call about a word choice in real time. AI can produce text that sounds reasoned. It cannot improvise a defence of choices it never actually made, because it didn't make them for reasons, it made them for probability.

This works because reasoning under mild pressure is very hard to fake convincingly if you didn't do the thinking. You'll stumble a bit. That's fine. Genuine explanation is allowed to stumble.

Don't argue with the score

Now, the bit people get wrong constantly: fighting the detector's number. Here's the thing about AI detection scores, and it's worth saying plainly, they're probabilistic guesses, not verdicts. The vendors who make these tools say so themselves, often in the small print nobody reads. False positives are documented and real; students and professionals alike have been flagged for writing that was entirely their own. Arguing "but the score says 80% AI" is arguing with a number that was never designed to be a courtroom finding in the first place. You can read more on how these tools actually perform, described plainly and without invented statistics, at scriptgrain.com/reference/ai-detector-accuracy.

So don't fight the number. Reframe the conversation. Say something like: "That score is a probability estimate, not a verdict, even the company that built it says that. Here's my process instead." Then hand over the drafts. That's a much stronger position than shouting about false positives in the abstract, because you're not disputing the tool, you're just declining to let it be the only evidence in the room.

Prevention is cheaper than defence

At the end of the day, the easiest way to win this argument is to never have to have it. Keep your drafts. Save your version history, don't clear it out the moment you submit something. And get a read on your own measurable voice before anyone questions it, not after. It costs nothing to run a quick analysis and understand your own patterns; it costs a lot more to reconstruct that understanding under accusation, flustered, with something at stake.

You didn't use AI. Fair enough, that's not really the debate anymore. The debate is whether you can show your work, and it turns out you always could. You just hadn't been asked before.