You wrote every word yourself, sent it off, and got back a message saying an AI detector flagged it. Now you're being asked to prove a negative, which is famously hard, and the client may be holding the invoice while you try. Here's how to prove you didn't use AI, and what to do if they still won't pay.
The short version: you can't prove a negative, so prove your process instead. A detector score isn't evidence of anything on its own: OpenAI withdrew its own detector because it wasn't accurate enough. What does convince clients is the trail real work leaves behind: version history showing the piece being written and rewritten, your notes and drafts, and your research. Reply calmly, in writing, with that evidence, and offer a call. If the client still refuses to pay, platforms like Upwork have a dispute process with deadlines you need to know before you need them.
Below: why detectors flag human writing, the evidence that actually works, a reply you can adapt, what to do if payment is on the line, and how to make sure it never happens again. The detector figures come from OpenAI's own published numbers and a Stanford study, checked in October 2026.
Photo by Vitaly Gariev on Pexels
Why AI detectors flag human writing (the false positive problem)
An AI detector doesn't know who wrote something. It estimates how predictable the text is and guesses from there. Clear, tidy, well-structured writing, the kind clients pay for, can look predictable to a model. That's why false positives happen, and why they happen to good writers.
The best evidence comes from the people who built one. When OpenAI launched its own AI classifier in January 2023, it reported that the tool correctly flagged only 26% of AI-written text and wrongly labelled human writing as AI 9% of the time. It called the tool "not fully reliable", said it was very unreliable on texts under 1,000 characters, and withdrew it in July 2023 because of its low accuracy.
Independent research is harsher. A Stanford study by Weixin Liang, James Zou and colleagues tested seven popular detectors on essays written by non-native English speakers. The detectors were near-perfect on essays by US eighth-graders, but flagged 61% of the non-native essays as AI on average, and 97.8% of them were flagged by at least one detector. If English isn't your first language, or you simply write in a plain, careful style, you're more likely to be flagged.
None of this stops clients using detectors. Gizmodo has reported on freelance writers who lost platform work after a detector flagged them, including writers who said they were let go by platforms that required them to use Grammarly. A detector score is a guess. Treat it as one, and help your client see it that way.
How to prove you didn't use AI: the evidence that works
Arguing with the score gets you nowhere. Showing how the work was made usually ends the conversation. In rough order of how convincing it is:
- Version history. If you wrote in Google Docs, File, then Version history, shows the document growing over hours or days, with timestamps. Microsoft Word keeps version history for files saved to OneDrive. A document that appears in one paste looks very different from one that was typed, cut and rewritten.
- Drafts and outlines. Early versions, a messy outline, the paragraph you deleted. Your scrappy first draft has finally found its purpose: alibi.
- Your research trail. Sources you read, notes, interview recordings or transcripts, and the questions you sent the client at the start.
- A live walkthrough. Offer a short call where you explain why you made the choices you made: the angle, what you cut, where a fact came from. Nobody can talk through a piece they didn't write the way its author can.
One thing not to lean on: running the piece through other detectors until one says "human". It can help to show the tools disagree, but a different guess doesn't prove much more than the first one did.
Photo by cottonbro studio on Pexels
What to say when a client accuses you of using AI
Wait before you reply. The first message you want to send is rarely the one that gets you paid. Then answer in writing, so there's a record, and keep it short: what happened, your evidence, and a way forward. Ask which tool they used and which passages it flagged, so you're responding to something specific.
Be honest about what you did use. If a grammar checker or AI tool touched the work at all, say so now and explain what it did. A small use you disclose costs you far less than a denial the client later finds out was only mostly true. Our guide on whether you have to tell clients you use AI covers what's expected.
If the client won't pay
Most accusations end once the client sees the evidence. When they don't, what you can do depends on where the work came from.
- Upwork, fixed-price. If a client ends a contract and asks for the funds back, you have seven calendar days to file a dispute, or the money is refunded to them. The client then has five days to respond; if they don't, the funds are released to you. If they decline, an Upwork mediator proposes a non-binding resolution, and after that the only option left is paid arbitration. On an active contract where the client simply won't release payment, Upwork says to contact support with proof the work is complete. The full steps are in Upwork's dispute guide.
- Fiverr. Fiverr's AI guidelines require you to honour a buyer's explicit request for non-AI work, and treat misrepresenting AI use as something that can end in a cancelled order, a refund and suspension. The stakes are higher, so keep your evidence ready and respond through the order page.
- Direct clients. Your contract and your invoice are what count. Point to the agreed scope, show the work was delivered, and offer a reasonable fix, such as a revision, before it turns into a standoff.
How to make sure it doesn't happen again
Prevention is mostly paperwork, done once:
- Write in a tool that keeps history. Draft in Google Docs or a OneDrive Word file from the first line, so the evidence exists before anyone asks for it.
- Put AI in writing up front. Say in your quote whether and how you use AI tools, and agree that a detector score alone isn't grounds to withhold payment. Our quote generator has a terms field for exactly that.
- Offer drafts along the way. Sharing an outline or a first draft mid-project builds trust and leaves a trail.
- Get paid in stages on big jobs. A deposit or milestones mean one accusation can't hold the whole fee.
Frequently asked questions
Can you be falsely accused of using AI?
Yes, and it happens often. OpenAI's own detector wrongly labelled human writing as AI 9% of the time before it was withdrawn, and a Stanford study found popular detectors flagged 61% of essays by non-native English writers as AI. Clear, careful writing is especially likely to be flagged.
How do I prove I didn't use AI?
Show your process rather than arguing with the score. Share the document's version history, your outline and drafts, and your research notes, and offer a short call to walk the client through your choices. Evidence of how the work was made is far more convincing than another detector result.
Are AI detectors accurate?
Not accurate enough to prove anything on their own. OpenAI withdrew its classifier in 2023 because of its low accuracy, after it caught only 26% of AI text. Detectors estimate how predictable writing is, which means polished human writing can be flagged and lightly edited AI text can pass.
What can I do if a client won't pay because they think I used AI?
Send your evidence in writing first. On Upwork fixed-price contracts, you have seven calendar days to file a dispute when a client ends the contract and requests a refund; mediation and then paid arbitration follow if you can't agree. With direct clients, your contract and invoice are what count.