Someone ran your writing through a detector, got a high number, and now you are being asked to explain yourself. You cannot disprove a probability, and trying to is the most common way people make their own position worse.
What you can do is shift the conversation from a score you did not produce to evidence you did. This article is the sequence that works, in order, and the moves that reliably backfire.
Before you reply
Answer the message, but do not improvise the substance of your defence in the first reply. A short acknowledgement — that you have received it, that you did not use AI in the way alleged, and that you will provide your drafts — buys you the time to assemble evidence properly.
Do not delete anything. Do not tidy up a document. Do not re-save a file before you have exported its version history. Editing evidence after an allegation looks worse than anything the detector reported, and it is the one mistake that is genuinely hard to recover from.
Above all, do not run the text through a humanizer or a paraphraser to lower the score. You would be altering the writing you are being asked to account for, and a lower score afterwards proves nothing about the original.
Ask for the evidence, specifically
A percentage on its own is not something anyone can evaluate. Ask, in writing and without hostility, for four concrete things:
- The passages. Which sentences or spans drove the result — not the document-level number.
- The tool and version. Which detector produced the report, and when.
- The published error rates. The false-positive rate the vendor states for that tool, and the conditions it was measured under.
- The procedure. What the institution’s or employer’s written policy says about how a detection result is used, and what your right of response is.
These are reasonable requests. Reviewers who have a real process will answer them. If a reviewer cannot say which passages were flagged, the conversation was never about evidence.
Gather process evidence
A detector looks at the finished text. Everything that surrounds the writing is evidence it structurally cannot see, and it is what actually persuades a fair reviewer:
- Version history. Google Docs, Word and most editors keep it. Export or screenshot it before anything else.
- Drafts and outlines. Earlier files, notes, mind maps, a plan scribbled in a notebook.
- Sources. The articles you read, your annotations, library or browser records, the PDFs on your drive.
- Timestamps. File creation and modification dates, commit history, submission logs.
- Earlier writing. Work in the same voice from before the assignment, ideally graded or already accepted.
- Correspondence. Questions you asked, feedback you received, messages about the work while it was in progress.
Volume is not the point. Two or three artefacts that show the writing developing over time do more than a folder of screenshots.
How to write the response
Keep it factual and short. State what you did, attach the evidence, and address the flagged passages one by one — where each idea came from, why you phrased it that way, what you were reading at the time. Specificity is what a percentage cannot answer.
If you used AI in a way your policy permits — brainstorming, an outline, grammar correction, translation — say so plainly and say exactly where. A disclosed, permitted use is a much better position than a partial account that unravels later.
You are not arguing that the detector is wrong in general. You are showing how this document was written.
It is legitimate to note, calmly and once, that detection scores are indicators rather than proof of authorship, that vendors publish false-positive rates, and that some institutions have disabled these tools for that reason. Say it once and return to your evidence. A long technical argument about detectors reads as deflection.
If English is not your first language
This is directly relevant and worth raising. Research on commercial GPT detectors found that more than half of TOEFL essays written by non-native English speakers were misclassified as AI-generated, while essays by US eighth-graders were classified almost perfectly. The signal being measured is linguistic range, not authorship.
You can state that as a documented finding rather than as a complaint, and cite it. The detail is in why AI detectors flag non-native English writers.
If the first conversation does not resolve it
Find the written procedure — academic integrity policy, student handbook, HR process — and follow the route it defines. Ask what the finding is based on, what the appeal path is, and what the deadlines are. Keep every exchange in writing, or send a short summary email after any verbal meeting.
Most institutions have a student union, ombudsperson, advisor or equivalent whose role is to support people through this process. Using them is normal and expected, not an escalation. Procedures vary widely between institutions and countries; the only reliable guide is the one that governs your case.
What does not work
- Detector shopping. Running the text through several free tools and sending the friendliest screenshot. Reviewers know tools disagree; a favourable result from an unnamed tool carries no weight.
- Rewriting the text. Any change after the allegation contaminates the evidence.
- A long essay on why AI detection is flawed. True in part, and it reads as avoidance. Two sentences, then evidence.
- Getting angry in writing. Your messages become part of the file.
- Admitting to something you did not do to end the stress quickly. An admission is usually final; the process is not.
The context you are entitled to
You are not being unreasonable in asking for more than a number. OpenAI withdrew its own AI text classifier in July 2023 for low accuracy, after publishing that it identified only 26% of AI-written text and flagged 9% of human text incorrectly. Vanderbilt University disabled Turnitin’s AI detector in August 2023, noting that a 1% false-positive rate across the roughly 75,000 papers it submitted in 2022 would mean around 750 wrongly flagged papers.
Detection tools have a legitimate use as a signal for review. IA Checker’s own position is the same one you should be asking for: a detector reports indicators and sentence-level evidence, and a person stays accountable for the decision. For the underlying numbers, see how accurate AI detectors really are, and for the short version of this page, what to do if you are wrongly accused.
This article describes a practical approach to an institutional process. It is not legal advice.