Flagged as AI but You Wrote It? What to Do if Falsely Accused

If you wrote it yourself and it was still flagged as AI, start from this fact: a detector score is a statistical prediction, and the companies that make the tools say so. Turnitin states its AI score should not be the sole basis for adverse action against a student, and GPTZero says it does not recommend using a detection result as the only proof for academic discipline. Your job is to show how the work was made.

The steps below work for a course assignment, a thesis chapter or a journal submission. They assume you wrote the text, or used AI only in ways your policy allowed.

First, do not touch the submitted file

The instinct is to rewrite the flagged passages or run the text through other detectors until one says "human". Resist it. The version that was flagged is the version under discussion, and a freshly reworded copy will look like a cover-up rather than evidence. Keep the original exactly as submitted, and keep your reaction off email until you have read the report.

1. PauseReply briefly that you wrote the work and would like to see the report and discuss it.
2. Ask for specificsWhich tool, which score, which sentences, and which policy the concern falls under.
3. Gather your processCollect version history, notes, outlines, sources and earlier drafts.
4. Understand the toolRead what the vendor itself says about thresholds, length and false positives.
5. Meet and explainWalk through how the piece developed and offer to discuss it in person.
6. Know the next stepAsk how decisions are made and whether there is a formal review or appeal route.

Ask exactly what was flagged

"Your paper is 40% AI" is not enough information to respond to. Ask for the report itself. In Turnitin, the AI indicator is visible only to instructors and administrators, but instructors can download the report and share it, according to Turnitin's AI writing FAQ. The highlighted sentences matter more than the headline figure.

When you have the report, look for patterns. Are the highlights in your introduction and conclusion? In stock methods language? In a short piece barely over 300 words? Each of those has a documented explanation, covered below and in our guide to the Turnitin AI detection score. If the highlights fall on your reference list or quotations, see why Turnitin flags references and citations as AI.

Gather evidence of your process

Nothing persuades like a visible writing history. Collect whatever you have:

• Version history. Google Docs and Microsoft 365 keep revision histories with timestamps. Export or screenshot the timeline showing the document growing over days.
• Earlier drafts and outlines. Files with different dates, comments from a supervisor, a draft you emailed to a friend.
• Research notes. Reading notes, annotated PDFs, a reference manager library with dates, lab notebooks, interview transcripts.
• Your sources. Be ready to explain why you chose each one and what it contributes.
• Any AI use you made. If you used a grammar tool or asked an assistant for feedback, say so plainly and show what it changed.

We have written more on this in how draft history can support academic integrity and keeping research notes that show your process.

Understand the limits of the tool

You do not need to argue that detectors are useless. You only need to show that a flag on its own is weak evidence, using the vendors' and researchers' own statements.

What vendors say. Turnitin reports a document-level false positive rate under 1% for documents with 20% or more AI writing, and in 2023 put its sentence-level rate at around 4%. It hides scores between 1% and 19% because false positives are more common there, and says that in documents of only a few hundred words the prediction is mostly all or nothing. It lists the kinds of human writing that are more likely to be misread: text without much structural variation, repetitive text, and paraphrase that does not develop new ideas. GPTZero describes its scores as probabilistic and gives an error rate of about 10% for its "moderately confident" label, per its guide to scan results. Sapling states that false positives and false negatives will occur.

What independent research found. Two studies come up most often.

Liang et al. (Patterns, 2023) ran 91 TOEFL essays by non-native English writers through seven detectors. The detectors were near perfect on US eighth-grade essays but misclassified the TOEFL essays at an average rate of 61.22%; 89 of 91 were flagged by at least one detector, and 18 were flagged by all seven. The authors linked this to lower perplexity in constrained, non-native prose. Our summary is in GPT detectors are biased against non-native writers.

Weber-Wulff et al. (International Journal for Educational Integrity, 2023) tested 14 tools, including Turnitin, and concluded that available tools were neither accurate nor reliable: all scored below 80% accuracy in their evaluation and only five scored over 70%. See our write-up of that study. Detectors have been updated since both papers, so cite them as evidence that false positives are a known, studied problem, not as a measure of today's error rate.

Have the conversation

Go in prepared and specific. Instructors and editors are usually looking for a reason to be confident in your authorship, and you can give them one.

• Lead with the work, not the tool. "Here is how this chapter developed" lands better than "detectors are unreliable".
• Walk through the timeline. Show the version history and point to where key paragraphs first appeared and how they changed.
• Explain your reasoning live. Offer to talk through your argument, sources and choices. Someone who built an argument can explain why it is shaped the way it is.
• Address the highlighted sentences. If they are generic transitions or a templated methods paragraph, say so and show where the template came from.
• Ask about process. Who decides, on what evidence, and what review or appeal route exists if you disagree.

Our guide on how to explain your AI writing process has wording for this, and instructors may find before questioning student authorship useful from the other side. Can professors detect AI? explains what else an instructor is likely to look at.

Journals and AI disclosure

For researchers, a flag from an editor is usually a question about disclosure as much as authorship. Publisher policies differ. Elsevier's policy for journals asks for a separate AI declaration naming the tool, its purpose and the extent of human oversight, but exempts basic grammar, spelling and punctuation checks. Nature Portfolio's AI policy sorts uses into risk tiers and treats language polishing as permitted, with disclosure framed as a matter of transparency and human accountability. If you used AI tools within the rules but did not declare them, the fastest fix is often an honest, specific declaration.

Preventing the next flag

Academic Humanizer can help before submission, not after an accusation. Its AI Detection Report shows which of your sentences read as AI to detectors, checked against Turnitin, GPTZero, ZeroGPT, Copyleaks, Sapling and Originality.ai, and the Humanizer rewrites only those sentences while keeping your argument, numbers and citations. Use it on drafts you are still revising, keep your version history as you go, and make sure any tool you use is allowed under your institution's or journal's policy. If formulaic phrasing keeps tripping detectors, why formal prose can look AI-written explains the pattern.