How often do AI detectors flag human writing as AI?

Why the published rates disagree

Each number answers a different question.

Who measured it. Vendors test their own detector, usually on the kind of text their customers submit, and set the threshold before they publish. Turnitin is open about the trade: to hold false positives near 1%, it accepts missing some AI text, so a document it scores at 50% could contain as much as 65% AI writing. Independent studies test several tools at once on sets their authors built, and those sets are often small.

What was tested. Liang and colleagues used TOEFL practice essays; Turnitin says they were all under 150 words and that its own detector refuses anything under 300. Weber-Wulff and colleagues wrote 54 documents, the human ones about 10,000 characters each, and ran them through 14 tools in spring 2023. A rate measured on one kind of writing tells you little about another.

Which level. A document-level rate counts whole papers wrongly flagged. Turnitin's sentence-level figure of about 4% means a highlighted sentence has roughly a 4% chance of being human-written, and Turnitin says just over half of those sentences sit right next to real AI text.

When. Detectors keep changing. Most independent numbers describe tools as they were in 2023, and Turnitin says it replaced its multi-model setup with a single model in July 2026.

What 1% means across a real course

A low rate still lands on real people. Take a course of 300 students who each hand in four essays they wrote themselves: 1,200 human essays. At a 1% document rate, about 12 would be flagged. At the 9% OpenAI measured for its own classifier, about 108. At the 61% average from the TOEFL study, most of them. This is arithmetic to show scale, not a measurement of any class.

Two things make it worse for some writers. Rates are averages, so if your writing is the kind detectors struggle with, your own chance is higher than the headline. And the fewer students who actually used AI, the larger the share of flags that land on honest work.

This is why Turnitin says its score should not be the sole basis for action, GPTZero says results should not be used to punish or as the final verdict, and OpenAI said its classifier should not be a primary decision-making tool.

Whose writing gets flagged more

The sources agree on the pattern even where their numbers differ.

Second-language writers. In Liang's study, the TOEFL essays that all seven detectors flagged used more predictable word choices. When the authors had ChatGPT enrich the vocabulary, the average false positive rate fell from 61.22% to 11.77%. When they simplified the vocabulary of the eighth-grade essays, it rose from 5.19% to 56.65%. Turnitin's own October 2023 test found no statistically significant gap between second-language and native writers for texts of 300 words or more, and a larger gap below that.

Plain, formulaic or concise writing. OpenAI's help center says its detector research pointed to this kind of writing being hit harder, and that one detector it trained labeled Shakespeare and the Declaration of Independence as AI-generated. Turnitin lists text with little structural variation, text that repeats itself and paraphrase that adds no new ideas.

Short texts. OpenAI called its classifier very unreliable below 1,000 characters. Turnitin needs at least 300 words of prose and says short documents get close to all-or-nothing predictions.

Machine-translated text. Weber-Wulff's team found accuracy on human writing dropped by about 20% once it had been machine-translated into English.

Autistic writers. A 2025 study of about 60,000 Reddit posts found one older open detector flagged under 2% of posts in each group, but significantly more from likely-autistic authors. One detector and one kind of text, so read it as an early signal.

What to do with a number someone shows you

If a teacher, editor or client quotes a detector result, ask four things: which tool, on what date, what the number means in that tool, and what your institution's policy says about scores. Turnitin's percentage, for example, is the share of qualifying prose its model thinks was AI-generated, not necessarily the share of the whole paper, and lists or bullet points are left out.

Then bring your process: version history, notes, an outline and earlier drafts. The page on what to do when a detector flags your essay covers that conversation, and the guide on proving you wrote your essay covers version history in Google Docs and Word.

Don't answer a flag by running your work through rewriting tools. It says nothing about how you wrote it, it can change your meaning, and Turnitin says its English detector also looks for text reworded by AI paraphrasers. If you use an editor on your own draft, HidenGPT included, keep the original beside the edited version, because no editing tool can tell you what a detector will report.