AI Detector False Positives: Why Human Writing Gets Flagged as AI
Human writing gets flagged as AI more often than you'd think. Here are the published numbers, who gets flagged most, and what to do about it.
- Free scan, no sign up
- Rhythm score
- Stock phrase check
- Runs in your browser
Forgot your code?
AI detector false positives happen when human writing is labeled as AI-generated. Detectors don't know who wrote a text. They score patterns, mainly how predictable the words are and how evenly the sentences are built, and a lot of real human writing has exactly those patterns. Plain vocabulary, even rhythm, formulaic formats and writing in a second language all raise the risk.
Quick answer: Published rates for AI detector false positives vary widely by tool and writer. OpenAI's retired classifier mislabeled human text 9% of the time. Turnitin reports under 1% per document for documents with at least 20% AI writing, and about 4% per sentence. A Stanford study saw over 61% of non-native English essays flagged. If you are flagged, don't admit to something you didn't do, show your version history and notes, offer to explain your reasoning and ask what the policy says.
How often detectors get it wrong: the published numbers
OpenAI's own classifier: 9% of human text mislabeled
OpenAI released an AI text classifier in January 2023. By its own figures, it labeled human-written text as AI-written 9% of the time and correctly caught only 26% of AI-written text. OpenAI took it down on July 20, 2023, citing its "low rate of accuracy" (OpenAI).
Turnitin: under 1% per document, about 4% per sentence
Turnitin says its document-level false positive rate is under 1% for documents with at least 20% AI writing, and its sentence-level rate is around 4%. It also found that 54% of falsely flagged sentences sit right next to actual AI writing, mostly in the transitions of mixed documents (Turnitin, June 2023).
Stanford: over 61% for non-native English writers
Researchers at Stanford ran 91 TOEFL essays by non-native English speakers through seven detectors. More than 61% were labeled AI-generated, and all seven detectors agreed on 19% of them. The same detectors scored US eighth-graders' essays almost perfectly (Liang et al., 2023).
The spread between these numbers is the point: a detector's accuracy depends heavily on who wrote the text and what kind of text it is.
Why human writing gets flagged
1. Plain, simple vocabulary
Detectors treat predictable word choice as a sign of AI. Clear, simple writing uses common words, so it scores as predictable. See what perplexity means in AI detection.
2. Even sentence rhythm
Sentences of similar length and structure look machine-made to a detector, even when a careful person wrote them. See what burstiness means.
3. Formulaic formats
Lab reports, legal writing, technical documentation, cover letters and five-paragraph essays follow templates. Templates are predictable by design.
4. Writing in a second language
People writing in a second language often use a smaller, safer vocabulary and standard sentence patterns, which is exactly what the Stanford study found detectors punish.
5. Heavy editing tools
Grammar checkers and rewriting tools smooth text toward the most standard phrasing. The more a tool polishes a draft, the more predictable it can look.
6. Short texts
A few sentences give a detector very little to go on, so scores swing more on short answers and paragraphs than on long documents.
7. Mixed human and AI text
When a document mixes your writing with AI-assisted parts, your sentences next to the AI parts are the ones most likely to be flagged, per Turnitin's own analysis.
What a score does and doesn't mean
A detector score is a probability based on patterns. It can't see your drafts, your notes or your screen, and it can't prove authorship. Two detectors often disagree on the same text. More on this in are AI detectors accurate?
What to do if your writing is flagged
- Don't admit to something you didn't do. A score is not evidence of how you wrote it.
- Show your process. Version history in Google Docs or Word, outlines, notes and sources show the work behind the text.
- Offer to explain it. Walking someone through your reasoning is strong evidence you wrote it.
- Ask what the policy says. Many schools and employers say detector scores alone aren't enough to act on.
More: what to do if you are wrongly flagged for AI.
How to check your own draft first
Paste your writing into the scanner above. Remove AI Tells isn't a detector and won't guess who wrote it. It shows the specific patterns detectors react to (stock phrases, AI vocabulary, punctuation habits and flat sentence rhythm) so you can see what might draw a flag and edit in your own words.
Frequently Asked Questions
What is an AI detector false positive?
It's when a detector labels human-written text as AI-generated. It happens because detectors measure writing patterns, not authorship.
How common are AI detector false positives?
It depends on the tool and the writer. Published figures range from under 1% per document (Turnitin's claim for documents with at least 20% AI writing) to 9% (OpenAI's retired classifier) to over 61% for non-native English essays in a Stanford study.
Why are non-native English writers flagged more often?
Writing in a second language often uses simpler vocabulary and standard sentence patterns, which detectors read as predictable, the same signal they use to spot AI.
Can editing tools make human writing look like AI?
They can. Grammar and rewriting tools push text toward standard, predictable phrasing, which is one of the patterns detectors look for.
Does Remove AI Tells detect AI writing?
No. It doesn't guess who wrote a text. It highlights the patterns that make writing read as machine-made so you can revise them.
Looking for other ideas? See all the ways to use Remove AI Tells.