← All notes

The accusation

Falsely accused of using AI

What a detector score actually means, why careful writers and non-native speakers get flagged most, and what to do in the days after the email arrives.

August 24, 20266 min readMichael Lynn

The email is short. Your work has been flagged. There is a number in it, usually a big round one, and a meeting on Thursday.

You wrote the thing. You remember writing it. And now you have to prove a negative to someone who has already seen a number that says otherwise.

Here is what that number is, what it is not, and what to do this week.

The number is not evidence of anything

An AI detector does not know who wrote your paper. It cannot. There is no watermark in ordinary prose and no residue a machine leaves behind that a person could not also produce.

What it measures is how closely your writing matches patterns that models produce often. Even sentence lengths. Certain words. A thesis at the top, three examples, a summary that restates the thesis. Those patterns are common in machine writing because that is what the training rewarded. They are also common in careful, formal, first-language-uncertain human writing, for reasons that have nothing to do with cheating.

So a high score says your prose has a shape. It does not say a model made it. No threshold turns the first thing into the second, and every vendor selling one knows it.

The people it flags are the people least able to argue back

This part is measured, and it has been for years.

Stanford researchers ran seven detectors over essays by non-native English speakers. The detectors called them machine-written 61% of the time. The same detectors were close to accurate on native writing. The reason is not mysterious. Someone writing in a second language reaches for the safe construction, the common word, the sentence pattern they trust. That is what the detector scores as machine-shaped.

Schools noticed. Vanderbilt turned Turnitin's AI detector off in 2023 and said plainly it could not tell them what they needed to know. Northwestern and the University of Texas did the same.

Courts have started to notice too. A student at Adelphi was accused on the strength of a Turnitin score his professor described as 100%, with nothing produced to back it up. He won. His record was cleared. Getting there cost his family over a hundred thousand dollars in legal fees. Another student lost a scholarship after using Grammarly to proofread.

None of that helps you on Thursday. But it means you are not the first person this has happened to, and the ground you are standing on is firmer than the email made it sound.

Do this before you reply

The single most useful thing you own is the history of the document, and it is easier to lose than you think.

Stop editing the file. Right now. Every save overwrites something.

Then go and collect what already exists:

  • Version history. Google Docs keeps it under File, then Version history, then See version history. Word keeps it in the cloud copy. This is the strongest thing you have, because it shows the piece being built over hours instead of appearing at once.
  • Drafts in your email. Anything you sent yourself, or a friend, or a study group.
  • Notes and outlines. Photographs of paper notes count. So do the tabs you had open, if your browser still has them.
  • The reading. Library records, PDFs with your highlights, anything with a timestamp before the draft.
  • Messages about the work. A text at eleven at night saying you are stuck on the third section is worth more than any argument you can make in a meeting.

Export or screenshot all of it. Put it somewhere that is not the same account as the file.

If you wrote it in one sitting with no history and no notes, say so plainly. That is a worse position and it is still a survivable one. Do not manufacture a paper trail. That turns a fight you can win into one you cannot.

How to write the reply

Short. Not angry. Not pleading.

State that you wrote it. Offer the evidence you have. Ask one question, and ask it in writing.

The question is: what is the policy, and what evidence supports the allegation beyond the detector score?

Ask it because you are entitled to know, and because it moves the conversation off the number and onto the process, where you have rights and the number does not. Most institutions have an academic integrity policy that says what counts as evidence and what your appeal route is. Find yours and read it before the meeting. It is usually a PDF nobody has opened.

Bring a person if your policy allows it. Take notes. Get decisions in writing.

If you are told the score alone is the case, that is the sentence to write down, because it is the weakest version of the argument against you and it will not survive an appeal in most policies.

Two things not to do

Do not run it through a rewriting tool and hand it back in. Even if you wrote every word. You have now submitted something a machine touched, in the middle of an integrity case, and you have turned a question about authorship into a question about honesty. It also does not work. Detectors retrain against those tools constantly.

Do not run your draft through a second detector and bring the result as proof. A low score is not evidence you wrote it, in the same way the high score was not evidence you did not. Bringing one in concedes that the machinery is a fair judge, which is the entire thing you are arguing against. If the other side puts a number on the table, the response is that the number cannot answer the question, and not a competing number.

If you did use a model

Then this is a different problem and the guide above is not for it.

The useful thing to know is that the cover-up is what ends careers, and the disclosure is what gets survived. Policies increasingly have a lane for assisted work that was declared. Almost none have a lane for assisted work that was denied and then proven.

Editing your own writing has never been cheating. Getting help with wording has never been cheating either, in most places, as long as the thinking is yours and you said so.

What a checker is actually good for

There is one honest use for these tools, and it is not defence.

It is looking at your own draft before you send it, so you can see which sentences are carrying the shape and change them if you want to. Not to score lower. To sound more like you.

That is what rewr.it/detect is for. It runs in your browser, it needs no account, and the words never leave the tab. It marks what it found and tells you which pattern each mark belongs to. We built it to check our own drafts before they went out under our own names.

We will not tell you it beats Turnitin. It does not, nobody's does, and you can read what we say about Turnitin in full. If what you want is your own writing back, without the phrasing that gets it mistaken for something else, that is the thing we can actually help with, and the cover letter is usually where people start.

The part worth holding onto

The number in that email is a reading of your prose. It is not a reading of you.

Bring your history, ask what the policy is, and make them say out loud what the case is besides a percentage. Most of the time there is not one.

You wrote it. Say so, once, and then make them do the work.

Say the thing you meant.
Just not the way a model would say it.

Paste a draft, see every tell named, and take the rewrite. No card to start.