This page is for people who wrote their own work and have been told a tool disagrees. It is not reassurance. It is the four things worth doing: understanding what the score you were shown actually measures, asking the institution a small number of specific questions, gathering the record while it still exists, and, if it helps, exporting a signed report of how you wrote it.
Two practical notes before anything else. Keep your files where they are and do not reorganise the folder, because the dates attached to them are part of what you have. And do not create anything now and present it as though you made it earlier. A gap you can explain is survivable. A document that turns out to have been made after the accusation is not.
We are not lawyers and nothing here is legal advice. It is a description of the technology and a checklist for the record you are assembling.
An AI detector is a statistical classifier. It reads a piece of text and estimates how closely the writing resembles the machine-generated text it was trained on. That estimate gets printed as a percentage, and the percentage then gets treated as a finding of fact. Those are different things, and the distance between them is where false accusations live.
It is: a classifier's estimate that your text resembles the machine-written examples it was trained on. That is a statement about style, not about who sat at the keyboard. No detector observed you writing, so none of them can report on it.
It is: a number produced by a statistical model that can be wrong, that is retrained without notice, and that frequently disagrees with other detectors given the same paragraph. A single percentage is a starting point for a conversation, not the end of one.
It is: usually an accuracy figure published by the company selling the detector, measured on text that company chose. The independent research points the other way. Stanford HAI researchers (Liang et al., 2023, arXiv 2304.02819) ran TOEFL essays written by non-native English speakers through seven detectors and found 61% falsely flagged as AI-generated. Every one of those essays had a human author.
It is: a score driven by general properties of the writing. Even sentence lengths, careful and correct grammar, formal register, textbook transitions, and a plain structure all push a score up. Those are the habits of second-language writers, of people who write to a template because school taught them one, of autistic writers, and of anyone using dictation or a grammar checker.
None of this means a detector is always wrong or that the person who ran it acted in bad faith. It means the output is evidence of a resemblance, that resemblance has a documented bias against particular groups of writers, and a percentage on its own does not establish who wrote a document.
Ask in writing, keep the tone flat, and ask for answers in writing. You are not arguing with the person who contacted you. You are finding out what the claim rests on, and most of these questions are ones an institution should be able to answer easily.
Detectors are retrained, and the same document can score differently on the same product a month apart. Without the tool name and the date, nobody can re-run the check, including the person accusing you. Ask for the full report rather than a screenshot of a number.
Most detectors return a percentage or a range, not a verdict. Somebody decided which number becomes an allegation. That decision is local policy, not a property of the tool, and it is a fair thing to ask about.
The relevant rate is not the headline accuracy figure. It is the rate for people who write the way you do: in a second language, with assistive software, after a brain injury, or simply in a plain, orderly style. If the institution does not have that figure for the tool it used, that is worth putting in writing.
This tells you what you are actually answering. If there is other material, such as a comparison with earlier work or a note from a reader, you need to see it. If the score is all there is, that is worth establishing early and calmly.
Ask for the procedure, the timeline, who makes the decision, what happens to your grade or your role while it runs, and whether you may bring someone with you. Getting this in writing costs nothing and keeps the process legible to you.
Do this before you reply to anything, because some of it disappears on its own. Version histories expire, browser history gets cleared, and file dates change when files move. Nothing on this list has to be impressive on its own. Taken together, an ordinary trail of a person working is difficult to fake and easy to recognise.
In Google Docs, File then Version history, which keeps named and timestamped revisions. In Word, check AutoSave and the OneDrive or SharePoint version list, and turn on Track Changes for anything you write from here on. This is usually the strongest thing an ordinary writer has, and it is also the easiest to lose by copying the file into a new document, so leave the original where it is.
Outlines, half-finished drafts, the version you emailed yourself at midnight, the copy on the USB stick. Do not tidy the folder. Created and modified dates are part of what makes the pile worth anything, and they change when files are moved between some systems.
Photographs of handwritten notes, annotated PDFs, library loans and downloads, the tabs still open from the night you researched it. Sources you can point to explain how your argument arrived at its shape.
Emails to your instructor or editor, questions asked in a class channel, a tutoring or writing-centre appointment, a message to a friend about the topic at a specific hour. These carry timestamps somebody else controls, which is exactly why they help.
Write down, for yourself, how you actually produced the piece: when you started, where you worked, what you read, what you cut, which parts were hard. Say plainly which tools you used, including a grammar checker, a translator, or a spell-checker, if you used them. A specific account you can repeat consistently is worth more than a general denial.
If you have more work due while this is unresolved, write it somewhere that records the history as you go. The editor here keeps a local trail of timings and sizes and can sign it. It also shows which of the 44 tells appear in your own writing, which is useful when you have to explain why plain, even prose is simply how you write.
Collect, do not construct. Retyping an old draft to make it look richer, backdating a file, or writing notes now and describing them as contemporaneous turns a false accusation into a real one. If part of your record is missing, say it is missing.
The editor can export a report about a document you are working on and sign it. The signature is ECDSA P-256 over SHA-256, produced with a key pair generated inside your browser by WebCrypto. The private key stays on your device. Anyone can check the signature at /verify, in their own browser, against the public key the report carries.
It is a sealed record, not a witness. The exact limits are set out below, so you can describe it accurately when you hand it over.
A signed report does not prove authorship. The key is generated in a browser, so a valid signature tells you that whoever held that key produced the report. It proves two narrow things: the report and the document it describes are unaltered since export, and any two reports carrying the same key fingerprint came from the same device. Describe it that way and it holds up.
It also only records from the moment you start using it. If the disputed work was written elsewhere, the trail for that document does not exist and cannot be reconstructed. What a report can do is turn a vague argument into a checkable artifact, and give the work you write from here on a history that somebody else can verify without taking your word for it.
If you learned English from textbooks, if you write in a formal and orderly register, if you use dictation or a grammar checker, or if you are rebuilding language after an injury, your writing has the properties these classifiers score as machine-made. That is a fact about the tools, not about your honesty.
of TOEFL essays written by non-native English speakers were falsely flagged as AI-generated across seven detectors. Every one of them was written by a person. Name the study precisely in any appeal: Liang et al., 2023, arXiv 2304.02819.
A checker that runs in your browser, with no upload step and no account, so you can read a paragraph without handing unsubmitted work to anyone. An explanation of every construction in your text that reads as machine-made, named and highlighted where it occurs. And a signed evidence report whose limits are printed above rather than buried, so you can hand it over and say exactly what it shows.
No tool can prove you wrote your essay, and nothing here is advice on your rights or your institution's disciplinary code. What the record and the checker do is replace a general denial with specifics somebody else can check.
The detectability score, the tell explanations and the evidence report are free, with no account. Nothing on this page is behind a paywall.