If you have been accused of using AI on work you wrote yourself, do three things before you reply. Save the record of how the piece was written, because version histories do not last forever and file dates change when a document is re-saved or copied. Read what the score you were shown actually claims, which is that your text resembles machine-written text, not that a machine wrote it. Then ask, in writing, which tool produced the result and what number your institution treats as an accusation. Most people do this in reverse. They reply within the hour, argue about a percentage, and go looking for drafts a week later when part of the trail is already gone.
Do not open the disputed document and start editing it. Depending on the app, saving can overwrite the most recent revision entry, and it moves the modified date to today. Today is the one date you do not want on that file.
Nothing on this list has to be impressive on its own. An ordinary trail of a person working across several days is hard to fake and easy to recognize. Gather it now, because some of it deletes itself on a schedule nobody announces.
An AI detector is a statistical classifier. It reads text and estimates how closely that text resembles the machine-generated writing it was trained on. That is a statement about style. No detector watched you write, so a score is not a finding of fact about authorship.
These models are retrained without notice, so the same document can score differently on the same product weeks apart, and two tools can disagree about the same paragraph. The clearest published measurement of how wrong they get it is Liang et al., Stanford HAI, 2023: across seven detectors, 61% of TOEFL essays written by non-native English speakers were falsely flagged as AI-generated. Every one was written by a person.
It is not random which writers get flagged. Even sentence lengths, correct grammar, a formal register, textbook transitions, and a plain reliable structure all push a score up. Those are the habits of second-language writers, of anyone taught to write to a template, of autistic writers, and of people using dictation. The mechanism is set out at why detectors flag human writing.
Ask in writing, keep the tone flat, and ask for the answers in writing. You are not arguing with the person who contacted you. You are finding out what the claim rests on.
You cannot prove a negative about a finished document. What you can show is how it came to exist. Roughly in order of weight: timestamps someone else controls, then a revision history of many small edits across days, then earlier drafts and outlines with their original dates, then notes and annotated sources, then your own account of how the piece was built.
One rule covers all of it. Collect, do not construct. Retyping an old draft to make it look fuller, backdating a file, or writing notes now and calling them contemporaneous turns a false accusation into a real one. If part of the record is missing, say so, because a gap you can explain is survivable. A fuller checklist, including what to do when there is no version history at all, is at you wrote it and were accused.
Keep the first message under a page. State plainly that you wrote the work, ask the questions above, list what you can provide, and ask for the written policy and the next step. Send it from your institutional address so it sits in the same record as everything else. Leave the argument about detector accuracy out of this first message, because it lands better once you know which tool ran and what threshold was applied.
In the meeting, the most persuasive thing in the room is detail about the work itself. Why that source and not the obvious one. What you cut and why. Which paragraph you rewrote three times and what was wrong with it. Somebody who did not write the piece cannot do that. Bring your record printed and ordered, say at the start if you used tools your course permits, and ask for the outcome and the reasoning in writing. We are not lawyers and this is not legal advice.
Rewriting the accused essay does not answer the question anyone asked you, and a copy edited after the accusation weakens the record you just spent an hour preserving. Leave that document alone until the process is finished.
What helps is knowing which of your own habits read as machine-made, so you can describe your style specifically instead of denying a score in general terms. The tell registry is free to read and names the exact constructions behind a machine-made reading, including the throat-clearing opener, the stacked hedges, and the transitions that join two sentences already in order. Running our checker on your own draft costs one credit out of ten free ones a day, with no account, no login, no card and no email address.
For the work you write from here on, keep a record a third party can check. The signed drafting record captures the history as you write and seals it with an ECDSA P-256 signature that anyone can verify without contacting us. The proofreader, word counter, readability checker, AI vocabulary checker and originality check are free and unlimited in the browser.
On policy, once: rewriting a draft you were permitted to generate is ordinary editing, and where AI is not permitted in submitted work, that rule covers the draft however it is edited. Read the policy that applies to you.
No. It estimates how closely your writing resembles the machine-generated text it was trained on, which is a statement about style rather than an observation of who wrote the document. These models are retrained without notice, they can disagree with each other on the same paragraph, and plain, correct, evenly paced prose is exactly what pushes a score up.
That is common and it is not fatal. Plenty of people write in one file and save over it. Fall back on the trail around the work: emails, messages, library loans, notes, annotated sources, and a specific written account of how the piece was built. Say plainly that the history is missing rather than assembling something that resembles one.
It is worth doing for one reason. The phrases named underneath the number tell you which of your own habits are being scored, which is what you need in order to explain your style in the meeting. Do not forward a second score as a rebuttal, because the question in front of the committee is authorship.
Say so early and plainly rather than letting it emerge later. Assistive software smooths text toward the phrasing a model considers likely, which is exactly the property these classifiers score, so it can raise a score without anyone doing anything wrong. Whether a given tool is permitted is set by your course policy, so read the one that covers you.
Not the one under review. Editing the disputed file after the accusation damages the record you are relying on, and it answers a question nobody asked. For future work, rewriting a draft you were permitted to generate is ordinary editing, and where AI is not permitted in submitted work, that rule covers the draft however it is edited.
Yes, and name it precisely: Liang et al., Stanford HAI, 2023, which found 61% of TOEFL essays by non-native English speakers falsely flagged as AI across seven detectors. It is short, it is specific about method, and a committee can look it up and check it.