Wrongly flagged by an AI detector? What to do next

If you have been wrongly flagged by an AI detector, start with this: a detector score is an estimate, not a finding. False positives are well documented, and the best answer to one is evidence of how you produced the work, such as drafts, version history, notes and your ability to talk about it. This guide covers what to do, in order, and how to prove you didn't use AI, whether you are a student or a professional writer.

AI detector false positives: a score is not proof

Every AI detector wrongly flags some human writing. The real questions are how often, and on what kinds of text.

• In a 2023 Stanford study, seven popular detectors flagged an average of 61.22% of TOEFL essays written by English learners as AI-generated.
• In 2025 the US Federal Trade Commission acted against a company whose detector was sold as 98% accurate. According to the FTC's complaint, independent testing found 53% accuracy on general-purpose content.
• Vanderbilt University switched off Turnitin's AI detector in August 2023. Turnitin had said at launch that its false positive rate was 1%; with 75,000 papers submitted the year before, Vanderbilt worked out that around 750 could have been wrongly labelled as partly AI-written.

The companies behind these tools say the same. Turnitin states that it "does not make a determination of misconduct" and that its AI percentage should not be the sole basis for action. OpenAI's guidance for educators says its own research didn't find detectors reliable enough for judgements with lasting consequences, and that a classifier it trained labelled Shakespeare and the Declaration of Independence as AI-written.

None of this means the person who flagged you acted in bad faith. Detector reports often look confident, with percentages and highlighted sentences. It does mean you have good grounds to ask for more than a score.

Why people get wrongly flagged by an AI detector

Knowing the common causes helps you explain your case calmly.

Plain or non-native English. Many detectors treat predictable word choices as a sign of AI. People writing in a second language, or writing simply and clearly on purpose, tend to use common words in common orders. The Stanford study traced its results to exactly this. We cover it in AI detectors and non-native English.

Repetitive or formulaic writing. Turnitin itself lists text without much structural variation, text that repeats itself and paraphrasing that doesn't develop new ideas as the human writing most likely to be mistaken for AI.

Short texts. A paragraph or a short answer gives a detector very little to go on. Ours refuses to give a result under 50 words and works best from 150.

Genres the tool wasn't tested on. The FTC alleged that the detector in its case had been trained only on academic content, though it was sold for everything. A false positive rate measured on one kind of writing says little about how a tool treats another.

Editing with other tools. In the same Stanford study, rewording essays with ChatGPT moved the detectors' results dramatically: simplifying the word choices in eighth-graders' essays raised the share flagged from 5.19% to 56.65%. Turnitin says ordinary grammar fixes from Grammarly were mostly not flagged in its tests, but text from Grammarly's generative features, such as drafting or paraphrasing, is likely to be. If your institution allows such tools and you used them, say so.

What to do, step by step

1. Stay calm and keep it factual. Don't agree to an account of events that isn't true, and don't rewrite or delete anything. Keep the work exactly as you submitted it.
2. Ask for the evidence in writing. Which tool was used, on what date, what exactly it reported, and which passages it highlighted. Ask which policy applies and whether it allows a decision based on a detector score.
3. Collect your version history, drafts and notes. The next section explains where to find them and what helps most.
4. Offer to talk through the work. Explain your argument, why you chose certain sources, what you cut. Offering to write a short piece on the same topic under supervision is a strong signal of good faith.
5. Point to the known limits of detectors. Briefly, with sources: the Stanford study, the FTC case, and the detector's own published error rates if it has any. The sources at the end of this page are a starting point.
6. Use the formal route if you need it. Ask about an appeal, and contact a student advocate, ombudsperson or union representative if your institution has one.

How to prove you didn't use AI: the evidence that helps

Nobody can prove a negative with certainty. What you can show is the work taking shape over time, and that is far stronger than any detector score. Gather what you have before any meeting.

• Version history in Google Docs. Click "Last edit" at the top right of the document to see earlier versions and who changed what. A reviewer needs edit permission to browse those versions, so share the file with edit access or make a copy for them. You can name up to 40 versions per document, which also stops Google merging them with others.
• Version history in Word. This works for files stored in OneDrive or SharePoint. In current Microsoft 365 apps, click the file's title and choose Version history; in Office 2016 to 2021, use File, then History. With a personal Microsoft account you can retrieve the last 25 versions.
• Separate drafts. Files saved at different stages, a first draft you emailed yourself, or a printout with your handwritten corrections.
• Notes and sources. Outlines, reading notes, highlighted PDFs, library loans and saved searches show the research behind the writing. Timestamps help.
• Writing-process records. Grammarly's Authorship feature, which Grammarly says is available on its free and Pro plans, records how a document was written in Google Docs, Word or the Grammarly editor and produces a report you decide whether to share. Some schools use Turnitin Clarity, which lets teachers replay a submission's drafting history.
• Your other work. Earlier assignments, especially anything written in class, show your normal voice and level.

Expect a reviewer to look at how the text grew. Many short sessions with rewrites and deleted sentences look like ordinary writing. A large block appearing all at once raises questions, even though there are innocent reasons for it. If you drafted in another app, on paper or in your phone's notes and pasted the result in, say so and bring the original. If you used an AI tool in a way that was allowed, say that up front too: being open about small things makes the rest of your account more credible.

In a meeting, treat the questions as a chance to show what you know. Turnitin's suggested opener for instructors is "Explain your process for completing this assignment," so expect something like it. Be ready to summarise your argument without notes, say where a source came from and what it says, walk through your version history, and discuss a highlighted paragraph: what you meant and what you changed.

Asking for the evidence: what a fair case includes

A fair process shows you what you are accused of and why. It is reasonable to ask:

• What did the detector actually report? A probability, a share of the text, a label, or a highlighted passage are different things.
• What is the tool's published false positive rate, and was it measured on writing like yours? Ask how much AI text it catches at the same setting too: a tool that rarely says AI about anything will also have a low false positive rate.
• Was anything other than the detector considered, such as your previous work or your drafts?
• Has the person reviewing the case read the text themselves, and what in it concerns them?

If the tool was Turnitin, three of its own rules are worth knowing. It needs at least 300 words of prose. It hides scores from 1% to 19% behind an asterisk, because false positives are more likely in that range. And instructors can download its AI report as a PDF, so you can ask to see it.

If the answer to the last two questions above is no, the case rests on a single score. Saying that out loud, without anger, often changes the conversation.

Should you run your work through other detectors?

It can help to show that tools disagree, but be careful about what it proves. A "human" result from another detector is weak evidence, because detectors miss AI text as well as flag human writing, so a clean result doesn't prove a person wrote something. Shopping around for the friendliest score doesn't help your case.

What a second detector can usefully show is that the evidence is shaky: different tools, different scores, and nothing settled. If you do check your work, consider privacy. Our detector runs entirely in your browser, so your text is never uploaded.

If you are a professional writer

Freelancers, journalists and authors get flagged too, by clients, editors or publishers. The same principles apply. Share drafts and version history, and offer a call to walk through your research and sources. If a client runs every piece through a detector, ask before the next job which tool they use, what result they treat as a problem, and what happens then. Agreeing on that in advance is much easier than arguing about it afterwards. There is more on this in our page for writers.

Protecting yourself next time

Write in a tool that keeps version history and leave it switched on. Name a version at each milestone, such as "first full draft", so it is easy to find later. Keep your notes and sources in one folder. If you are allowed to use a grammar checker or other AI help, note what you used. Ten seconds of record-keeping now can save hours of stress later.

Sources

1. Liang et al., GPT detectors are biased against non-native English writers (Patterns, 2023), checked October 2026
2. FTC: FTC Order Requires Workado to Back Up Artificial Intelligence Detection Claims (28 April 2025), checked October 2026
3. Vanderbilt University: Guidance on AI Detection and Why We're Disabling Turnitin's AI Detector (16 August 2023), checked October 2026
4. Turnitin Guides: Turnitin's AI writing detection capabilities FAQs, checked October 2026
5. Turnitin Guides: What questions do I ask my student if AI is detected?, checked October 2026
6. OpenAI Help Center: How can educators respond to students presenting AI-generated content as their own?, checked October 2026
7. Google Docs Editors Help: Find what's changed in a file, checked October 2026
8. Microsoft Support: View previous versions of Office files, checked October 2026
9. Grammarly: Authorship, checked October 2026
10. Turnitin: Turnitin Clarity, bringing transparency to the student writing process (4 March 2025), checked October 2026