AI DetectionStatisticsAI DetectionFalse Positives

AI Detection Statistics (2026): Accuracy, False Positives and Use

Verified AI detection statistics for 2026: detector accuracy, false positive rates, bias against non-native writers, and how students and teachers use AI.

By The Humanizer Team8 min read
Cover illustration for AI Detection Statistics (2026): Accuracy, False Positives and Use

Key takeaways

  • A 2023 study of 14 detection tools found none reached 80% accuracy, and its authors called the tools neither accurate nor reliable.
  • Stanford researchers found seven detectors classified an average of 61.22% of TOEFL essays by non-native English speakers as AI.
  • Turnitin claims a document-level false positive rate under 1%, while Bloomberg found GPTZero and Copyleaks falsely flagged 1% to 2% of human essays.
  • Use keeps climbing: HEPI found 92% of UK undergraduates used AI in 2025, and a CDT survey found 68% of US teachers used AI detectors.
  • Many figures come from vendors or older tool versions, so check the source and date before relying on any number.
In this guide
  1. AI detection statistics at a glance
  2. How accurate are AI detectors?
  3. How common are false positives?
  4. Who gets falsely flagged most often?
  5. How much AI do detectors find in student work?
  6. How many students use AI?
  7. How many teachers and schools use AI detectors?
  8. What do these AI detection statistics mean for you?
  9. How did we choose these statistics?
  10. The bottom line

AI detectors are less reliable than their marketing suggests. Independent studies keep finding errors in both directions, with human writing flagged as AI and edited AI text passing as human. Non-native English writers get flagged most often. Meanwhile, use keeps rising on both sides: more students use AI tools, and more teachers run detectors.

Below are the AI detection statistics we could verify, grouped by theme. Every figure links to its source, whether that's a peer-reviewed study, an official page or a news outlet's test. When a number comes from a company describing its own product, we say so.

AI detection statistics at a glance

StatisticFigureSource
TOEFL essays by non-native English speakers flagged as AI, average across seven detectors61.22%Stanford HAI
Best accuracy among 14 detection tools testedBelow 80%Weber-Wulff et al., 2023
Average detector accuracy on unedited AI text, then after simple manipulation39.5%, then 22.14%Perkins et al., 2024
AI text correctly identified by OpenAI's AI Classifier before it was withdrawn26%Decrypt
Turnitin's claimed document-level false positive rateUnder 1%Turnitin
Turnitin's sentence-level false positive rate, as reported in 2023About 4%Turnitin
Pre-ChatGPT essays falsely flagged by GPTZero and Copyleaks1% to 2%Bloomberg
ZeroGPT's AI score for the Declaration of Independence97.93%Decrypt
AI-written exam answers that markers didn't detect94%PLOS ONE
UK undergraduates using AI in some form in 202592%HEPI
US middle and high school teachers using AI detection tools68%Education Week

How accurate are AI detectors?

Accuracy depends on what's tested: raw AI output, edited AI text or human writing. Independent results are weaker than vendor claims.

  • No tool reached 80%. Researchers led by Debora Weber-Wulff tested 14 tools, including Turnitin, and found that "even the highest accuracy values are below 80%" (Weber-Wulff et al., 2023). They concluded the tools "are neither accurate nor reliable" and leaned toward labeling text as human.
  • Light editing cut detection. In the same study, AI text that a person edited scored 42% on the authors' accuracy measure, compared with 74% for unedited AI text (Weber-Wulff et al., 2023).
  • Simple tricks made it worse. A 2024 study found detectors identified unedited AI text correctly only 39.5% of the time on average, and got human-written control samples right just 67% of the time (Perkins et al., 2024). After techniques such as adding spelling errors, average accuracy fell to 22.14% (Perkins et al., 2024).
  • OpenAI gave up on its own detector. OpenAI's AI Classifier correctly flagged 26% of AI-written text and wrongly labeled 9% of human text as AI in OpenAI's evaluation (Decrypt). OpenAI withdrew it in July 2023, citing its "low rate of accuracy."
  • Results swing between tests. In a 2024 Durham University study of 60 short physics essays, ZeroGPT reached 98% accuracy and QuillBot 83% (Yeadon et al., 2024). Months later, Decrypt found ZeroGPT rated the Declaration of Independence 97.93% AI (Decrypt).

Vendor figures are a separate category. ZeroGPT advertises 98.4% accuracy on its homepage, and GPTZero says its "highly confident" results have an error rate under 2% (GPTZero). They may hold on the companies' own test sets, but we found no independent test that reproduces them. For more on individual tools, see is GPTZero accurate and is ZeroGPT accurate.

How common are false positives?

A false positive is human writing labeled as AI. It's the error that hurts students most, and the rate depends on who's counting and what text is tested.

  • Turnitin's claim: under 1% for documents with more than 20% AI writing, which Turnitin says means it "might flag a human-written document as AI-written for one out of every 100 fully-human written documents" (Turnitin FAQ).
  • How Turnitin tests it: before every model update, Turnitin says it tests on more than 700,000 academic papers written before ChatGPT was released (Turnitin FAQ).
  • Sentence level: in June 2023, Turnitin said its sentence-level false positive rate was "around 4%," and that 54% of falsely flagged sentences sat right next to actual AI writing (Turnitin).
  • Low scores are least reliable: Turnitin hides AI scores from 1% to 19% because its testing found "a higher incidence of false positives" in that range (Turnitin).
  • What 1% means at scale: Vanderbilt University submitted 75,000 papers to Turnitin in 2022 and estimated that a 1% false positive rate could have wrongly flagged around 750 of them (Vanderbilt).
  • Real essays: Bloomberg ran 500 college application essays written before ChatGPT existed through GPTZero and Copyleaks, and the tools falsely flagged 1% to 2% as likely AI (Bloomberg).
  • Lightly polished text: a 2025 study of 12 detectors found they "frequently flag even minimally polished text as AI-generated" (Saha and Feizi, 2025).

We explain why these errors happen in why AI detectors have false positives and how AI detectors work.

Who gets falsely flagged most often?

Second-language writers. It's the best-documented bias in AI detection.

  • Non-native English speakers: in a Stanford study of seven detectors, an average of 61.22% of TOEFL essays written by non-native English speakers were classified as AI-generated (Stanford HAI).
  • Near-unanimous errors: all seven detectors flagged 18 of the 91 TOEFL essays (19%), and 89 of the 91 (97%) were flagged by at least one (Stanford HAI). The same detectors were "near-perfect" on essays by US-born eighth-graders.
  • Machine-translated writing: accuracy on human-written English texts was 96% in the Weber-Wulff study, but it dropped by 20% for human texts written in other languages and machine-translated into English (Weber-Wulff et al., 2023).
  • Round-trip translation: Originality.ai's own 2026 study found human writing translated into Spanish and back into English was falsely flagged 28.02% of the time, and 27.84% for Portuguese, while direct translations stayed low (Originality.ai).

If English is your second language, our page for ESL writers explains what helps, and our guide to AI detectors in other languages covers non-English writing.

How much AI do detectors find in student work?

  • Turnitin's first year: by March 21, 2024, Turnitin had reviewed over 200 million papers. Over 22 million, about 11%, had at least 20% AI writing, and over six million, about 3%, had at least 80% (Turnitin). These are what the detector flagged, not confirmed cases.
  • Human markers miss most AI work: in a University of Reading test, 94% of AI-written exam submissions went undetected, and they earned grades half a grade boundary higher than real students on average (Scarfe et al., 2024).
  • Proven cases are rising: a Guardian investigation found almost 7,000 proven cases of AI cheating at UK universities in the 2023 to 2024 academic year, or 5.1 per 1,000 students, up from 1.6 per 1,000 the year before, based on data from 131 universities (eWeek).

How many students use AI?

  • UK undergraduates: 92% used AI in some form in 2025, up from 66% in 2024, and 88% had used generative AI for assessments, up from 53% (HEPI).
  • AI text in submitted work: 18% of students in the same survey said they had included AI-generated text directly in their work (HEPI Policy Note 61).
  • US teens: 26% said they had used ChatGPT for schoolwork in a fall 2024 survey of 1,391 teens, up from 13% in 2023 (Pew Research Center).
  • What teens think is fair: 54% said it's acceptable to use ChatGPT to research new topics, but only 18% said the same about writing essays (Pew Research Center).

How many teachers and schools use AI detectors?

  • US teachers: 68% of middle and high school teachers said they used AI detection tools, in a Center for Democracy and Technology survey of 460 teachers fielded in late 2023 (Education Week). That was a 30 percentage point jump from the year before (K12 Dive).
  • Pullbacks: Vanderbilt disabled Turnitin's AI detector in August 2023, citing the false positive math above and concerns about transparency (Vanderbilt).
  • Turnitin's own guidance: the company says its percentage "should not be used as the sole basis for action or a definitive grading measure by instructors" (Turnitin FAQ).

What do these AI detection statistics mean for you?

Before relying on one tool, check how it reports results. Our comparison of the best AI detectors explains each score, and what AI percentage is acceptable covers thresholds. For a quick look at your own draft, our free AI detector estimates AI-likeness from writing patterns. Treat it as a pre-check, not a verdict.

How did we choose these statistics?

We included only figures we could trace to a study, an official company page or a reputable news test, and each one links to its source. Company claims about their own products are labeled. Dates matter: many tests used 2023 or 2024 versions of detectors and AI models, and both have changed since. When you cite a number, cite its date too.

The bottom line

The research points the same way: AI detectors make mistakes in both directions, edited AI text slips through, and non-native English writers are flagged far more often. Meanwhile, AI use and detector use keep climbing. Treat any score as a starting point, and keep evidence of how you wrote.

If you're polishing an allowed AI-assisted draft or stiff writing of your own, try The Humanizer free with no credit card. It's designed to read naturally, though results vary by text.

Frequently asked

Questions, answered

Written by

The Humanizer Team

Editorial

Ready to humanize your AI content?

Paste text from ChatGPT, Claude or Gemini and get natural writing designed to pass GPTZero, Turnitin and other AI detectors.

See plans & pricing
  • 350 free words every month
  • No credit card required
  • Sign in with Google in one click

Keep reading