AI Detector False Positives
What the peer-reviewed research says about how often AI detectors are wrong, why non-native English writers are flagged most, and what to do in the first 48 hours.
Being Flagged For Work You Wrote
You wrote the essay. The report says it is AI-generated. There is no appeal button, the score comes with no explanation, and the people assessing it often believe the number is more reliable than it is.
The Case You Can Actually Make
False positives are measured, published and substantial, and the gap falls hardest on students writing in a second language. Knowing the actual figures, preserving the right evidence and moving inside your institution's appeal window are what decide these cases.
How it works
- Do not edit the flagged work or resubmit it
- Preserve drafts, notes, outlines and version history immediately
- Ask for the full report, not the percentage
- Put your account in writing and request the formal process if needed
- Check your institution's appeal deadline — missing it is how students lose
What the research actually shows
The most cited peer-reviewed study on this is Liang, Yuksekgonul, Mao, Wu and Zou, GPT detectors are biased against non-native English writers, published in Patterns in July 2023. The researchers ran seven widely used GPT detectors over 91 TOEFL essays written by non-native English speakers and 88 essays by native-speaking US eighth-graders.
| Finding | Result |
|---|---|
| Average false-positive rate, non-native English essays | 61.3% |
| Essays flagged unanimously by all seven detectors | 19.8% |
| Performance on native-speaker essays | Near-perfect |
| False-positive rate after improving word choice | Fell to 11.6% |
| Effect of simplifying native-speaker vocabulary | Flagging increased |
The last two rows are the important ones. The same essays, with more sophisticated vocabulary, stopped being flagged. Native-speaker essays, with simpler vocabulary, started being flagged. The detectors are substantially measuring how elaborate the writing is — not who or what wrote it.
Citing this in an appeal
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. and Zou, J. (2023) 'GPT detectors are biased against non-native English writers', Patterns, 4(7), 100779. doi: 10.1016/j.patter.2023.100779. It is open access, peer-reviewed and from Stanford — which makes it considerably harder to wave away than a blog post.
Who gets flagged most
False positives are not randomly distributed. If you are in one of these groups, your risk is structurally higher and it is worth knowing before you are accused rather than after.
- Students writing in a second language. The measured gap is enormous — 61.3% against near-zero. This is the dominant factor.
- International students on taught masters programmes, who combine second-language writing with an unfamiliar academic register.
- Writers who edit heavily. Repeated revision moves prose toward the conventional, which is exactly what detectors score.
- Anyone writing formulaic sections — methods, definitions, literature summaries — where the phrasing options are genuinely limited.
- Neurodivergent writers whose drafting produces unusually regular sentence structure.
- Students using grammar tools such as Grammarly, which smooth text in the direction detectors read as machine-written.
The first 48 hours
What you do immediately matters more than what you argue later, because most of it is about evidence that becomes harder to produce with time.
- Do not touch the document. Do not rewrite the flagged passages, do not re-run it through anything, do not resubmit. Altering work after an allegation is far more damaging than the original flag.
- Preserve everything now. Drafts, notes, outlines, photographs of handwritten planning, browser and library history, the emails where you asked your supervisor a question. Export it rather than trusting it to stay available.
- Check your version history. Google Docs keeps it under File → Version history; Word Online and OneDrive keep it too. This is usually the single strongest piece of evidence, because it shows the document being built.
- Request the full report. Ask for the sentence-level breakdown, not the percentage. You cannot respond to a number, and the breakdown often shows the score concentrated in exactly the generic passages you would expect.
- Put your account in writing before any meeting — how you researched it, when you wrote it, what sources you used. Written first, so it cannot be characterised afterwards.
- Find the appeal deadline. It is in your academic integrity policy or student handbook. Missing the window is the most common way students lose cases they would have won.
What to argue, and what not to
Version history showing the document being built over time. The peer-reviewed false-positive rates. Turnitin's own guidance that the score is a conversation starter, not a finding. Offering to discuss the content, or to write something comparable under observation.
Running the text through a humanizer and offering the result. Deleting or editing files. Arguing that detectors are "always wrong". Missing the deadline and appealing afterwards. Agreeing to an informal resolution before you understand what it records.
The strongest position is narrow and factual: the tool produces false positives at a published, substantial, peer-reviewed rate, that rate is highest for writers in your situation, and here is the documentary record of this specific document being written. Avoid arguing that detection is worthless in general — it invites a debate you do not need to win.
If you are offered an oral examination, take it
Many institutions offer a viva or portfolio review as an alternative form of verification. If you wrote the work, this is the fastest route to resolution, and declining it tends to be read the wrong way.
Reducing the risk next time
None of this is about writing to please software. It happens that the things which lower false-positive risk are the same things that earn marks.
- Draft in Google Docs or Word Online. Version history accrues automatically and costs nothing. This is the single most useful habit.
- Keep your outline and notes in the same folder as the draft, dated.
- Write with specifics. Named sources, actual findings, your own examples. Generic prose is what flags, and it is also what loses marks.
- Check before you submit so a high score is something you have already looked at rather than something that arrives as an accusation.
- Know your institution's policy on AI tools and on detection evidence. Some have restricted detector use entirely; that is worth knowing in advance.
Want the method, not just the tool?
How to Humanize ChatGPT Text (Without Making It Worse) — Why AI text reads the way it does, how detectors actually work, why they flag innocent students — and the rewriting moves that make writing genuinely yours.
Try it on your own document
Write, cite, humanize and check your AI score in one editor. The free plan includes limited AI chat and AI detection each day, plus three humanizer runs — no card needed.
Open the editor →