AI detector false positive rates: the numbers, with sources

AI-detection scores are now used in admissions, coursework and hiring, but the evidence on how often they are wrong is scattered across papers and company blog posts. This page collects the most-cited figures in one place, each linked to its original source, so you can quote them accurately — in an appeal, a conversation with your instructor or your own research.

The numbers at a glance

Non-native writers are flagged far more often

The most important study for ESL writers is Liang, Yuksekgonul, Mao, Wu and Zou, published in Patterns in 2023. The researchers ran 91 TOEFL essays written by human test-takers through seven widely used AI detectors. On average the detectors labeled 61.22% of them as AI-generated, and 89 of the 91 essays were flagged by at least one detector. Essays written by US eighth-graders were classified almost perfectly (Liang et al., 2023).

The authors traced the gap to perplexity — how predictable each word is to a language model. Non-native writers tend to use more common vocabulary and simpler structures, which lowers perplexity and looks machine-like. GPTZero, one of the best-known detectors, describes perplexity and burstiness as indicators it uses (GPTZero).

Turnitin's own figures

When Turnitin launched its AI indicator in 2023 it promoted a document-level false positive rate below 1%. A few weeks later its chief product officer published more detail: the under-1% figure applies to documents the tool scores above 20% AI; the sentence-level false positive rate is around 4%; and 54% of false-positive sentences sit right next to genuinely AI-written ones. Scores below 20% are now shown with an asterisk because they are less reliable (Turnitin, 2023).

Vanderbilt University disabled the feature, noting that even a 1% rate could have meant about 750 of the 75,000 papers it submitted in 2022 being wrongly flagged (Vanderbilt, 2023). For more, see what a Turnitin AI score means.

Independent tests: no tool above 80%

A team of academic-integrity researchers led by Debora Weber-Wulff tested 14 detection tools, including Turnitin and GPTZero. All scored below 80% accuracy, and only five scored above 70%. The tools leaned towards classifying text as human, and their accuracy dropped further when AI text was paraphrased or machine-translated. The authors concluded the tools are neither accurate nor reliable enough to be used as proof (Weber-Wulff et al., 2023).

OpenAI's withdrawn classifier

OpenAI, the company behind ChatGPT, released its own AI-text classifier in January 2023. It correctly identified only 26% of AI-written text and mislabeled human text as AI 9% of the time. OpenAI warned it “should not be used as a primary decision-making tool” and withdrew it on July 20, 2023 because of its low accuracy (OpenAI, 2023).

What the numbers mean for you

• A score is not proof. Every source above — including the companies that build detectors — treats the result as an estimate.
• Your process is your best evidence. Version history, drafts and notes show how the work was written. See what to do if you're accused.
• Quote the research accurately. Use the figures and links on this page in an appeal letter.
• Natural English helps. Fixing translated-sounding phrasing in your own writing removes some of the patterns detectors associate with AI — without changing what you say.