Skip to content
Sound More Human

Blog

If an AI detector flagged your own writing: an evidence checklist

· 6 min read

By

You wrote the paper, and a detector says a large share of it was written by a machine. The first hours matter: what you do now decides how much you can show later. This is a checklist for that situation. It assumes the flag is wrong. If you used a tool in a way your course does not allow, none of this will help, and the honest move is to talk to your instructor or your institution’s integrity office.

What the flag is, and what it is not

A detector gives a probability-style score, not a finding. The vendors say so themselves. In a March 2023 post, Turnitin’s chief product officer wrote that the company does not make a determination of misconduct (opens in a new tab), that its false positive rate is not zero, and that instructors must apply professional judgment and knowledge of their students. The same post advises instructors to assume positive intent when the evidence is unclear.

The independent evidence is less comfortable for the tools:

  • A Stanford team tested seven detectors on 91 essays that non-native English speakers wrote for the TOEFL exam, and on 88 essays by US eighth-graders. The detectors were close to accurate on the eighth-graders. On the TOEFL essays, the average false-positive rate was 61.3 per cent, and 97.8 per cent of them were flagged by at least one detector. Every essay was human-written. Liang et al., Patterns, 2023 (opens in a new tab). The authors say themselves that the samples were small, which is why the finding is a warning and not a rate you can quote for any particular student.
  • Weber-Wulff and seven co-authors tested 14 detection tools, including Turnitin, and concluded they are neither accurate nor reliable (opens in a new tab). They also found that human writing machine-translated into English lost accuracy by about 20 per cent.
  • Vanderbilt University turned off Turnitin’s AI detector (opens in a new tab) in August 2023. It noted that a 1 per cent false-positive rate, applied to the 75,000 papers it submitted in 2022, would mean around 750 papers wrongly labelled.

None of this proves your paper is human-written. It is the reason a score alone should not settle your case, and you are entitled to say so, politely, with the sources.

Before you reply

  • Do not edit the flagged document. Do not rewrite passages, do not re-save over the file, and do not resubmit a changed version. The record you need is the one that exists now.
  • Do not delete anything: drafts, notes, browser tabs, chat threads with classmates, search records.
  • Reply in writing, calmly, and ask for a meeting. Angry first emails get quoted back.

The evidence checklist

  1. Ask exactly what you are accused of and on what basis. Which tool and version, which passages, what score, and whether anything besides the score supports it. Ask for the report itself, in writing. A concern you cannot see is one you cannot answer.
  2. Pull the version history. In Google Docs, choose Last edit at the top right (opens in a new tab) to see who changed the file and when. In Word, files stored in OneDrive or SharePoint have a Version history panel (opens in a new tab); Microsoft says a personal account can retrieve the last 25 versions, so do this promptly. A document that grew across many sessions over days is good evidence. It is not proof, and you should never build or backfill history afterwards: it can be checked, and doing so would be dishonest.
  3. Gather everything that is not the final text. Outlines, notes, highlighted PDFs, photographed handwriting, earlier drafts, library or database searches, emails or messages where you talked through the idea.
  4. Show your sources. Be ready to say where each citation came from and what you took from it. A paper whose sources you have actually read is hard to fake.
  5. Be ready to explain it. The strongest evidence in most cases is that you can talk about your paper: what your argument is, why you chose this example over another, what you cut, what you would say to an objection. Practise aloud beforehand.
  6. Offer other samples of your writing. Earlier graded work, in-class writing or emails. Vanderbilt’s guidance to its own instructors suggests comparing a suspect paper with the style, tone and level of the student’s earlier work, so having it ready helps.
  7. Say what tools you used, truthfully. A spell checker, a translator, a grammar tool, a citation manager. If your course requires disclosure, say what you disclosed and when. Hiding a small allowed use looks worse than mentioning it.

In the meeting

Bring the timeline, the files and a short written summary of how you wrote the paper. Answer the questions you are asked, and say when you do not remember something. If the conversation turns formal, ask what the process is, who decides, and what your deadlines are for responding or appealing. Write the answers down.

A first reply you can adapt (written for this article)

Dear [name],

Thank you for letting me know. I wrote this paper myself and would like to show you how. Could you send me the report you are relying on, including the tool used and the passages it flagged? I can bring my drafts, notes and the document’s version history, and I am glad to talk through the argument and my sources in person. When would be a good time to meet?

[Your name, course and section]

If it goes further

Find your institution’s written procedure and read it for deadlines. Ask whether there is a student advocate, an ombudsperson or a writing centre that can sit in or help you prepare. Keep every message. In the formal write-up, state what the flag rests on, cite the research above, and put your evidence in date order. Ask that the decision rest on more than a score. The tool’s own maker says instructors must apply judgment.

If English is not your first language

The Stanford finding is about you specifically, and you are entitled to say so. Everything above applies, and this article on second-language writing goes into the evidence and how to keep your process visible. If your first draft was in another language and you translated it, keep that draft. It is strong evidence, and your course may require you to mention the translation.

Where this leaves a rewriting tool

We cannot say what any detector would report about any text, before or after a rewrite, and we would distrust anyone who says they can. If you are allowed to use a tool like ours, keep the same records you would keep for any other help: your draft, the rewrite, and a note of what you kept. For the policy side of that question, see how to read your course’s AI policy. The home page tool keeps your original beside the rewrite so the comparison is always there.