Falsely accused of using AI? Here is what actually works.

Last updated 23 September 2026

Someone believes you did not write your own work. A client is withholding an invoice, an editor has gone quiet, a publication has pulled something, or a committee has written to you. You know you wrote it. That turns out not to be the same thing as being able to show it.

This guide is what we know after building tools for exactly this situation. It is written to be useful whether or not you ever use our product. Some of it is uncomfortable, including the parts about what evidence cannot prove.

1. Start with what is actually happening

There are two very different situations that feel identical from the inside, and they need different responses.

Somebody ran a detector and it flagged you. This is the one people assume they are in. It is less common than you would think. Independent testing of the three best-known tools found that on genuinely human writing, two of them flagged nothing at all out of 495 passages, and the third flagged 3.8%. Modern detectors are much better at avoiding false positives than their reputation suggests.

Somebody formed a judgement and is looking for a reason. This is the more common case, and it is the one that costs people money. A reader notices you used an em dash, or three-item lists, or wrote a confident sentence, and concludes a machine did it. No software was involved. The accusation is social, and the detector gets invoked afterwards as justification.

Why the distinction matters: in the first case you are arguing with a score. In the second you are arguing with a person's impression, and evidence that reads as defensive makes it worse. What works in both cases is the same thing, delivered calmly: a documented process record.

2. Four things that make it worse

Do not add typos or scuff up your writing

This has become the standard advice and it is bad advice on three counts. It does not work against a determined accuser. It makes your writing worse, permanently. And it is a tell in itself: a paragraph that is flawless except for one "accidental" typo reads as manufactured.

Do not lead with how inaccurate the detector is

It is true, and it is the wrong opening. You are asking a busy person to accept that their tool is broken. They will defend the tool. Lead with what you can show about your own work instead, and let the tool's unreliability be a footnote rather than your argument.

Do not delete the mess

Your outlines, half-finished drafts, notes to yourself, the WhatsApp message where you complained about paragraph four, the version you saved at 2am. People tidy these up before a dispute because they are embarrassed by them. They are the most persuasive thing you own.

Do not accept a deadline you cannot meet

Accusations often arrive with pressure attached. If someone wants an answer within 24 hours, ask for a week, in writing. Assembling a real record takes a day or two, and a rushed denial is weaker than a considered response.

3. The evidence that exists, ranked

Ordered by how much weight it carries with a sceptical third party, not by how easy it is to produce.

4. What no service can honestly promise you

• That a particular client, editor or committee will accept the evidence. People decide, not software. No honest vendor can promise an outcome.
• That you can prove a negative retroactively. If your writing went through a tool that discarded the process record, the record is gone. This is why setting up evidence before you need it matters so much.
• A percentage. Any service offering "98% human" is guessing, and a confident-sounding number that later collapses does you more damage than no number at all.
• Proof of authorship. It does not exist as a technical artefact. What exists is evidence of process, and a person willing to weigh it.

5. If a contract says "must pass AI detector"

These clauses are spreading through freelance and agency contracts, and they are worth understanding precisely, because two similar-sounding clauses put risk in completely different places.

"No AI used" is a promise about your process. You control it. You can satisfy it by not using AI.

"Must pass AI detector" is an outcome warranty depending on someone else's software, a threshold they may not have defined, and a model that may change without notice. A writer who did every single word themselves can still fail it. It hands your payment to a tool you do not control.

The practical rule: never accept a clause that makes payment depend solely on a detector score. If a client insists on AI language in the contract, make it about process, disclosure, documentation and reasonable review — all things you can actually evidence.

6. If a client is withholding payment

1. Get it in writing. Ask, politely, which tool produced which result, and what threshold was applied. Detector scores are frequently not reproducible, and this question often ends the matter.
2. Send a short, factual response with your evidence attached, and keep it to one page. Long defensive emails read as guilt.
3. Point at the contract. If the agreement says "original work" rather than "must pass a detector", the detector result is not a contractual condition, whatever the client believes.
4. Record the pattern. If this client has done it before, that is useful context for any later dispute.
5. Know your escalation. For small amounts, a platform dispute process or a small-claims filing may be more practical than a lawyer. For larger sums, a letter before action is cheap and often decisive.

7. If it is a school or university

Academic processes are different, and the most important difference is that you usually have a right to a process — to see the evidence, to respond, and to appeal. Nearly every detector vendor's own documentation says their output should not be used as the sole basis for a misconduct finding. Turnitin says so about its own AI writing report. GPTZero's guidance says its results should not be used to punish students.

What helps: the version history of the document, your drafts, your notes, and a written statement describing how you worked. What does not help: refusing to engage, or arguing that the detector is racist or broken. Both may be true and neither is your defence. Your defence is your process.

8. If it is a publisher, agent or platform

Publishing disputes move fast and reputational damage is the real cost. The most useful thing you can do is get ahead of it with a clear, unemotional account plus documentation, before a decision is made without you. If a publisher has already acted, note that contracts increasingly contain AI warranty clauses, and the Authors Guild publishes model clauses and guidance that is worth reading before you respond to anything.

9. Language you can copy

Short, factual, unemotional. Adapt as needed.

The honest summary

You cannot prove a negative, and anyone who tells you otherwise is selling something. What you can do is show how the work was made, from a record you did not create, and let a reasonable person draw the obvious conclusion. That is a much stronger position than a denial, and it is available to almost everyone who actually did the work — usually in a place they have not thought to look.