A false positive is a detector calling human writing AI-written. It is the error that gets students accused, so it is the number to look at first. Below: what each detector maker says about its own rate, then what independent researchers measured, then why the two often seem to disagree. Every figure is dated, because most detectors shipped new models between July and October 2026.
What the detector makers say
| Detector | Its own false-positive figure | Where and when |
|---|---|---|
| Turnitin | Under 1%, for documents with over 20% AI writing. Scores from 1% to 19% show only as an asterisk, with no highlights. | FAQ, updated Oct 14, 2026; the asterisk rule since Jul 16, 2024 |
| GPTZero | "No more than 1%"; its GPTZero 4o model flagged 1 of 10,402 essays in its own test | Technology page (undated); 4o launch post, Oct 18, 2026 |
| Pangram | 0.0041% (41 of 1,000,000 web texts); 0.019% on English academic writing | Pangram 4 model card, Jul 29, 2026 |
| Copyleaks | 0.30% at its default sensitivity (0.16% to 0.66% across its three settings) | V11 testing methodology, Sep 28, 2026 |
| Winston AI | 7 false positives in 6,000 documents (99.88% of human writing recognized) | Winston 5.0 launch post, Sep 24, 2026 |
| Originality.ai | 99.954% of human essays recognized on the PERSUADE 2.0 set (n = 25,996) | Originality 5.0 post, Oct 20, 2026 |
| ZeroGPT | "<1%" | Homepage (undated) |
| Grammarly, QuillBot, Sapling, Scribbr | No false-positive rate published. Sapling says false positives fell 60% in internal tests and are likelier on short, generic, essay-like text. | Each detector's own page, read Oct 22, 2026 |
These are each company's own measurements on data it chose. Two things to know when you read them. First, "99% accurate" or "#1 on RAID" usually refers to how much AI text a tool catches after its threshold has been set to flag a fixed 1% or 5% of human text; it isn't a measured false-positive rate. Second, a rate measured on long, clean English essays won't hold on a 150-word paragraph.
What independent tests found
| Study | Human writing tested | False positives found |
|---|---|---|
| Harvard and NYU researchers (Isley, Gaebler & Goel), preprint, Sep 2026 | Master's-programme application essays from 2020-2022, before ChatGPT | GPTZero "AI, high confidence": 0.1% to 0.3% of essays per year. Counting "mixed" labels at any confidence: 1.0% to 1.6%. Pangram "AI": 0.0% to 0.1%; with "mixed": 1.4% to 2.8%. |
| Epoch AI, Jul 2026 | 495 passages by published authors, pre-2022 (blogs, fiction, science) | Pangram 0 of 495; GPTZero 0 of 495 |
| Vrije Universiteit Brussel, Int. J. for Educational Integrity, Jun 2026 (peer-reviewed) | 40 pre-2019 theses by non-native English-speaking graduate students | Turnitin, GPTZero, Copyleaks and Pangram all put all 40 in the 0-20% band |
| Notre Dame, preprint, Aug 2026 | 306 research abstracts from 2013-2015 | Pangram 0%, GPTZero 0% |
| Graphite, May 2026 | 15,700 web articles published 2020-2022 | GPTZero 1.355%, Pangram 1.844%, counting "AI-assisted" and "mixed" as AI |
| Chung-Ang University, preprint, Aug 2026 | 104,752 pre-ChatGPT manuscripts by non-native English writers | 13 free open-source detectors only: from 0.0% (Binoculars) to 100.0% (Log-rank); RoBERTa 93.1% |
| Chicago Booth (Jabarian & Imas), NBER working paper, Sep 2025 | 1,992 pre-2020 passages (reviews, blogs, news, novels, résumés) | At a 0.5 threshold: GPTZero 0.71%, Pangram 0.00%; the open-source RoBERTa detector 95.03% |
| Pratama, PeerJ Computer Science, Jun 2025 (peer-reviewed) | 72 journal abstracts published by 2021 | GPTZero 0%, ZeroGPT 16.67%, DetectGPT 31.94%; 44.44% flagged by at least one of the three |
| Bloomberg Businessweek, Oct 2024 | 500 Texas A&M application essays from summer 2022 | The two detectors it tested falsely flagged 1% to 2% |
| RAID benchmark, ACL 2024 (peer-reviewed) | 14,971 human-written documents | Commercial detectors at a 0.5 threshold: 0.00% (GPTZero) to 1.42% (ZeroGPT). Open-source ones up to 99.3% (GLTR) |
| Perkins et al., 2024 | 10 human-written samples, 7 tools | Turnitin flagged none of the 10; across all seven tools the authors say "only 67% of the tests were accurate" on the human samples |
| Weber-Wulff et al., IJEI, Dec 2023 (peer-reviewed) | 9 human-written documents, 14 tools (tested Mar-Apr 2023) | "Ranged from 0% (Turnitin) to 50% (GPT Zero)" |
| Liang et al. (Stanford), Patterns, Jul 2023 (peer-reviewed) | 91 TOEFL essays by non-native writers; 88 US 8th-grade essays | 7 detectors (March 2023 versions) flagged 61.3% of the TOEFL essays on average, and 97.8% were flagged by at least one. The US essays were classified "near-perfect". |
| OpenAI's own classifier, Jan 2023 | Its "challenge set" of English texts | 9%; OpenAI withdrew the tool on Jul 20, 2023, citing its low accuracy |
The pattern: on long English prose, today's commercial detectors flag well under 2% of human writing in most independent tests, and several 2026 tests found none. Older detector versions (2023), free open-source checkers and short texts are where the large numbers come from.
Why the same detector can have a 0.1% or a 2% false-positive rate
1. Which labels count as "AI". Most detectors don't just say AI or human. GPTZero gives "mixed" results and confidence levels, and Pangram also reports mixed or AI-assisted text. The 2026 Harvard/NYU study found that counting "mixed" labels and any confidence level raised the essay-level false-positive rate "by an order of magnitude". If your school treats a "mixed" result as a flag, the vendor's headline rate doesn't apply to you.
2. From one essay to one student. A 1% rate per essay adds up over a degree. In the same study, when any one of an applicant's essays counted, the share of pre-ChatGPT applicants flagged at least once reached 6.8% for GPTZero and 11.9% for Pangram in 2022, counting "AI" plus "mixed" labels at any confidence. Vanderbilt made the same point in 2023: 75,000 papers a year at a 1% rate is about 750 papers.
3. Length and type of text. Every rate above was measured on passages of hundreds of words. Turnitin raised its minimum from 150 to 300 words in May 2023, "based on our data and testing"; Sapling says short, generic and essay-like text is likelier to be flagged; the Booth study found rates up to 3% on shorter passages against 1% or less on medium and long ones.
4. Non-native English writers. The widely quoted 61% comes from the 2023 Liang study of seven detectors as they were in March 2023. Later tests of the current commercial tools found far lower rates on non-native writing: Turnitin's own 2023 test gave 1.4% on 2,221 second-language documents against 1.3% on first-language ones; the 2026 VUB study found no flags on 40 non-native theses. The Chung-Ang study shows free open-source checkers still flagging most non-native writing.
5. Free open-source checkers. Several free sites run open-source models. In the RAID benchmark, the Booth paper and the Chung-Ang study, some of these flagged 90% or more of human writing. A "100% AI" result from an unnamed free checker means little.
Turnitin's false positives
Turnitin is the detector most universities provide: of the 53 universities in our census of 300 that provide a detector, 44 use Turnitin's. Students can't run it, so its rate comes from Turnitin and a few outside tests:
- Turnitin's claim: under 1% false positives for documents with over 20% AI writing, checked on more than 700,000 pre-ChatGPT papers before each model release. It hides scores from 1% to 19% behind an asterisk because false positives are more common in that range.
- Sentence level: Turnitin said in May and June 2023 that about 4% of the sentences it highlights may be human-written, most often right next to AI-written text. It hasn't published a sentence-level figure for any newer model, including the English model it updated on Oct 14, 2026.
- Outside tests: Temple University's teaching centre found 28 of 30 human samples scored correctly, with one rated 11% AI and one not rated; Perkins et al. found 0 of 10 flagged; Weber-Wulff et al. found 0%; the 2026 VUB study found none of 40 non-native theses flagged. The first three used small samples and 2023 versions.
More on what Turnitin does and doesn't detect: Does Turnitin detect AI humanizers?
If a detector flagged your own writing
- The detector makers say a score isn't proof. Turnitin's FAQ says its AI indicator "should not be used as the sole basis for action". GPTZero says the same of its results.
- Many universities say so too. MIT's 2026 report says its Committee on Discipline doesn't treat detector output alone as sufficient evidence; Liberty University says results alone are not "an indictment"; Arizona State's College of Liberal Arts and Sciences calls a result a starting point for a conversation. Look up what your school's own page says (247 universities, 48 with a written answer).
- Bring your process, not another score. Drafts, notes and the version history of your document show how the text was written. Running it through a second detector proves little either way, given how much free tools disagree.
Sources (read Oct 22-23, 2026)
- Turnitin, "Turnitin's AI writing detection capabilities FAQs" (updated Oct 14, 2026) and "AI writing detection model" release notes: guides.turnitin.com
- Turnitin blog, "AI writing detection update from Turnitin's Chief Product Officer", May 23, 2023, and "Understanding the false positive rate for sentences", Jun 14, 2023: turnitin.com/blog
- Turnitin blog, "New research: Turnitin's AI detector shows no statistically significant bias against English Language Learners", Oct 26, 2023: turnitin.com/blog
- GPTZero, technology page (undated) and "Introducing GPTZero 4o", Oct 18, 2026: gptzero.me
- Pangram, Pangram 4 model card, Jul 29, 2026: pangram.com/research/model-card/pangram-4
- Copyleaks, "AI detector testing methodology" (V11), Sep 28, 2026: copyleaks.com/ai-detector/testing-methodology
- Winston AI, "Introducing Winston 5.0", Sep 24, 2026: gowinston.ai
- Originality.ai, "Introducing Originality 5.0 and Max", Oct 20, 2026: originality.ai/blog
- ZeroGPT homepage; Sapling, Grammarly, QuillBot and Scribbr detector pages (all undated, read Oct 22, 2026)
- Isley, Gaebler & Goel, arXiv 2609.22549, Sep 18, 2026
- Epoch AI, "AI detectors' false negatives" data insight, Jul 15, 2026: epoch.ai
- Van Vlasselaer, Van Droogenbroeck & Spruyt, International Journal for Educational Integrity 22:16, Jun 29, 2026
- Karr, Khvatskii, Hua & Chawla, arXiv 2608.11256, Aug 2026
- Paredes, Druck, Benson & Smith (Graphite), "AI now writes as many online articles as humans do", May 15, 2026: graphite.io
- Park, Jeong & Kim, arXiv 2608.26710, Aug 27, 2026
- Jabarian & Imas, "Artificial Writing and Automated Detection", NBER Working Paper 34223, Sep 2025
- Pratama, PeerJ Computer Science 11:e2953, Jun 23, 2025
- Davalos & Yin, "Do AI Detectors Work? Students Face False Cheating Accusations", Bloomberg Businessweek, Oct 18, 2024
- Dugan et al., "RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors", ACL 2024, arXiv 2405.07940
- Perkins et al., arXiv 2403.19148, Mar 2024 (International Journal of Educational Technology in Higher Education 21:53)
- Salem, Fiore, Kelly & Brock, "Evaluating the Effectiveness of Turnitin's AI Writing Indicator Model", Temple University (undated)
- Weber-Wulff et al., "Testing of detection tools for AI-generated text", International Journal for Educational Integrity 19:26, Dec 25, 2023
- Liang, Yuksekgonul, Mao, Wu & Zou, "GPT detectors are biased against non-native English writers", Patterns 4(7), Jul 2023
- OpenAI, "New AI classifier for indicating AI-written text", Jan 31, 2023 (withdrawn Jul 20, 2023)
- Vanderbilt University, "Guidance on AI detection and why we're disabling Turnitin's AI detector", Aug 16, 2023