Which AI detectors claim a false positive rate under 1%, and who checked each claim?
Nine of the 72 AI detector vendors in the index claim a false positive rate under 1% on their own pages, by the vendor claim field read on 8 and 9 November 2026: Turnitin (under 1%), Copyleaks (0.30%), GPTZero (1 in 10,402), Pangram (1 in 10,000), Winston AI (zero across 27,743 documents), Originality.ai (99.954% specificity, a rate of 0.046%), ZeroGPT (under 1%), Proofademic (0.2%) and Undetectable AI (0%). Seven of the nine have been tested by at least one study in the index, and the study figures run from 0.0000 of 1,992 human passages for Pangram in Jabarian and Imas 2025 (NBER w34223, read 9 November 2026) to 52 of 114 human submissions for GPTZero in Orenstrakh 2023 (arXiv 2307.07411, read 9 November 2026); Proofademic and Undetectable AI have no indexed study behind the claim. Codequiry's 1.3% is the only vendor rate in the index above 1%, and it is stated as a limit rather than a claim.
A claim and a check are different numbers from different hands. The claim column quotes the vendor's page with its date; the check column lists every false positive figure an indexed study prints for the same tool, with the base, so a reader can see where a 1 in 10,000 and a 9 in 18 came from. The claim vs evidence gap ranking does the same for accuracy claims, and the false positive rate ranking sorts the study figures alone. This page is one of the questions answered from the index; its siblings ask whether a 100% accurate detector exists and whether detectors work on machine-translated text, and the AI detector register carries every card.
| Detector | Vendor claim, as published | Claim source and read date | Who checked, and what they found | Study records |
|---|---|---|---|---|
| Turnitin | "under 1% for documents with over 20% of AI writing", tested on "over 700,000 additional academic papers" before each release | guides.turnitin.com, 8 November 2026 | Walters 2023: 0 of 42; Weber-Wulff 2023: 0 of 18; Perkins 2024: 0 of 10; Van Vlasselaer 2026: 0 of 40 | 5 (one is Turnitin's own statements) |
| Copyleaks | 0.30% at the Balanced default (methodology page); .03% on the product page | copyleaks.com, 8 November 2026 | Walters 2023: 0 of 42; Elkhatat 2023: 1 of 5; Perkins 2024: 5 of 10; Orenstrakh 2023: 1 of 114 English, 10 of 10 Spanish; Van Vlasselaer 2026: 0 of 40 | 6 |
| GPTZero | "only incorrectly predicting 1% of human texts as AI" (FAQ); "1 in 10,402" on student writing in the GPTZero 4o post of 18 October 2026; 0.08% on GPTZero's own four-domain benchmark sets for model 4.3b, vendor-measured, benchmarking page of 5 February 2026 updated 20 July 2026 (vendor benchmarks record) | gptzero.me/faq, 8 November 2026 | RAID 2024: 0.00% at threshold 0.5; Jabarian and Imas 2025: 0.0071 of 1,992; Walters 2023: 2% of 42; Perkins 2024: 1 of 10; Elkhatat 2023: 1 of 5; Weber-Wulff 2023: 9 of 18; Orenstrakh 2023: 52 of 114; Van Vlasselaer 2026: 0 of 40; Epoch AI 2026: 0 of 495 pre-2022 passages | 16 |
| Pangram | "1 in 10,000 False positive rate", "calculated on an aggregate of public datasets containing tens of millions of written documents"; 70 of 781,981 (0.0090%) on the Pangram 4.1 model card of 19 October 2026, vendor-measured (vendor benchmarks record) | pangram.com, 8 November 2026 | Jabarian and Imas 2025: 0.0000 of 1,992 at threshold 0.5, at most 0.0075 per genre at Youden thresholds; Van Vlasselaer 2026: 0 of 40; Epoch AI 2026: 0 of 495 pre-2022 passages | 4 |
| Winston AI | "zero false positives across 27,743 human documents in published evaluations" | gowinston.ai, 8 November 2026 | Weber-Wulff 2023: 0 of 18; RAID 2024: 0.55% at threshold 0.5, 0.75% at 0.25 | 2 |
| Originality.ai | Human specificity 99.954% (5.0) and 99.958% (5.0 Max) on PERSUADE 2.0, a false positive rate of 0.046% and 0.042% by subtraction | originality.ai, 8 November 2026 | Walters 2023: 0 of 42 (2 uncertain), model 1.4; RAID 2024: 0.25% at threshold 0.5; Jabarian and Imas 2025: 0.0011 of 1,992; Orenstrakh 2023: 7 of 114; Epoch AI 2026: 19 of 495 (3.84%), 16 of them fiction | 6 |
| ZeroGPT | "<1% False positive rate" (product metrics); the home FAQ states an error rate lower than 2% | zerogpt.com, 8 November 2026 | Weber-Wulff 2023: 0 of 18; Perkins 2024: 0 of 10; Walters 2023: 2% of 42; RAID 2024: 1.42% at threshold 0.5 | 5 |
| Proofademic | "0.2% document false-positive rate" on "a held out set of 4,200 fully human authored academic texts", March 2026 | proofademic.ai, 8 November 2026 | No indexed study has tested the tool | 0 |
| Undetectable AI | "100% detection / 0% false positives", the vendor's summary of a journal study of ten free detectors, in the comparison table on the detector page | undetectable.ai, 9 November 2026 | No indexed study has tested the tool; the journal study the vendor cites is not yet an index record | 0 |
Claims as quoted on each card from the vendor page read on the date shown; study figures as each paper prints them, with the base, from the records linked in the sections below. GPTZero's FAQ figure of 1% is at the line rather than under it; the row is included for the 1 in 10,402 figure in the vendor's 18 October 2026 post.
Which claims did an outside study check?
Seven of the nine, with between 2 and 11 records each. The checks agree with the claim in some rows and not in others, and the base decides how much either number means. Pangram's 1 in 10,000 is matched by the one large-sample test: Jabarian and Imas 2025 recorded 0.0000 false positives on 1,992 human passages at a 0.5 threshold and at most 0.0075 per genre at Youden-optimised thresholds, and Van Vlasselaer 2026 recorded 0 of 40 human papers (NBER w34223; DOI 10.1007/s40979-026-00226-w, read 9 November 2026). Turnitin's "under 1%" is consistent with four zeros on bases of 42, 18, 10 and 40 (Walters 2023, Weber-Wulff 2023, Perkins 2024, Van Vlasselaer 2026), though the author of Walters 2023 notes the papers may have been in Turnitin's training data.
GPTZero's claims meet the widest spread of checks. RAID 2024 recorded 0.00% at threshold 0.5 on 14,971 human documents, and Jabarian and Imas 2025 recorded 0.0071 of 1,992 at every fixed threshold, both under 1% (arXiv 2405.07940, read 8 November 2026; NBER w34223, read 9 November 2026); Weber-Wulff 2023 recorded 9 of 18 (50%) and Orenstrakh 2023 recorded 52 of 114 (45.6%), both on student or translated writing (DOI 10.1007/s40979-023-00146-z, read 8 November 2026; arXiv 2307.07411, read 9 November 2026). Copyleaks' 0.30% stands beside 0 of 42 and 0 of 40 in two studies and 5 of 10 and 10 of 10 Spanish papers in two others (Perkins 2024; Orenstrakh 2023). Winston AI's zero across 27,743 documents meets 0 of 18 in Weber-Wulff 2023 and 0.55% in RAID 2024 (arXiv 2405.07940, read 8 November 2026). ZeroGPT's under 1% meets two zeros, a 2% and RAID's 1.42% (false positive rate ranking).
On 9 November 2026 the index added Epoch AI 2026, which scored 495 human passages written before 30 November 2022 by 99 professional authors: Pangram 3.3.2 and GPTZero model 2026-05-11-base each flagged 0 of 495 (0.00%, 90% confidence interval 0.00% to 0.54%) and Originality.ai Turbo 3.0.2 flagged 19 (3.84%), 16 of them fiction, at each detector's default verdict (epoch.ai, results table, read 9 November 2026; the insight is not peer reviewed). The same day the index recorded the two vendors' own current figures as vendor measurements, dated and never ranked: GPTZero's benchmarking page states a 0.08% false positive rate for model 4.3b averaged over four domains of 1,000 human texts each (page of 5 February 2026, updated 20 July 2026, gptzero.me, read 9 November 2026), and Pangram's 4.1 model card states 70 false positives in 781,981 pre-ChatGPT English texts, 0.0090% (19 October 2026, pangram.com, read 9 November 2026); both sit in the vendor benchmarks record.
Which claims has nobody checked?
Two. Proofademic's 0.2% rests on a held-out set of 4,200 human-authored academic texts in a March 2026 evaluation the vendor ran itself (proofademic.ai, read 8 November 2026), and no study in the index has tested the tool. Undetectable AI's 0% is the vendor's summary of a study of ten free detectors in the Indian Journal of Psychological Medicine, shown in a comparison table on the detector page without a base (undetectable.ai, read 9 November 2026); the index has not read that paper, so the figure is filed as a vendor statement until it is. Both cards sit in the vendor table of the false positive rate ranking, unranked, under the same rule that kept Pangram's 1 in 10,000 there until Jabarian and Imas 2025 was indexed on 9 November 2026.
Why does a checked claim still differ from the check?
Two questions added 9 November 2026 follow from this one: the published test data question asks which of these vendors name or size the set their figure came from, and the false positive studies question counts how many outside records print a human-text figure for each card.
Because the texts differ. Every vendor figure in the table was measured on data the vendor chose, and every study figure on data an outside author chose, and the two agree where the texts resemble each other and diverge where they do not. The low study figures sit on large general corpora: RAID's 14,971 documents and the 1,992 pre-2020 passages of Jabarian and Imas 2025 (DOI 10.18653/v1/2024.acl-long.674, read 8 November 2026). The high figures sit on student coursework, non-native and machine-translated writing: GPTZero's 52 of 114 in Orenstrakh 2023, its 9 of 18 in Weber-Wulff 2023 with 6 on translated texts, and Liang 2023's average of 61.22% across seven detectors on 91 TOEFL essays (DOI 10.1016/j.patter.2023.100779, read 8 November 2026). A vendor's 0.046% on PERSUADE essays and a study's 7 of 114 on databases coursework can both be exact.
Base size matters as much as text type. A zero on 10, 18 or 42 documents cannot rule out a rate of several percent; only RAID 2024, Jabarian and Imas 2025 and Orenstrakh 2023 used more than 100 human texts (annual false positive evidence review). A claim of 1 in 10,000 cannot be checked on 42 papers at all; it can only fail to be contradicted.
No detector score should be the only evidence in a misconduct, hiring or publishing decision, and a false positive rate under 1% still means one wrongly flagged writer in every hundred to ten thousand.
Questions about false positive claims
Which AI detector claims the lowest false positive rate?
Winston AI and Undetectable AI both state zero, Winston AI across 27,743 human documents in published evaluations (gowinston.ai, read 8 November 2026) and Undetectable AI with no base (undetectable.ai, read 9 November 2026). Among rates with a stated base, Pangram's 1 in 10,000 (0.01%) is the lowest, and it is the one matched by a large outside sample, 0.0000 of 1,992 in Jabarian and Imas 2025.
Does GPTZero claim under 1%?
The FAQ states 1% exactly, on RAID, and the GPTZero 4o post of 18 October 2026 states 1 in 10,402 on diverse student writing corpora (gptzero.me/faq, read 8 November 2026). Outside studies recorded 0.00% at threshold 0.5 in RAID 2024 and 9 of 18 in Weber-Wulff 2023 (Weber-Wulff 2023 record).
Has any study confirmed a sub-1% false positive rate?
On large samples, yes for three tools: Pangram at 0.0000, Originality.ai at 0.0011 and GPTZero at 0.0071 of 1,992 human passages in Jabarian and Imas 2025 (NBER w34223, read 9 November 2026), and GPTZero at 0.00%, Originality.ai at 0.25% and Winston AI at 0.55% at threshold 0.5 in RAID 2024 (arXiv 2405.07940, read 8 November 2026). The same tools posted higher rates on student and translated text in other studies. Epoch AI 2026 added a third large base on 9 November 2026: 0 of 495 pre-2022 human passages for Pangram and GPTZero, and 19 of 495 (3.84%) for Originality.ai (epoch.ai, read 9 November 2026).
Which vendor claim has no outside test at all?
Proofademic's 0.2% and Undetectable AI's 0%, as of 9 November 2026; neither tool appears in any of the 35 indexed records.
Is Codequiry's 1.3% a claim under 1%?
No, and it is the only vendor rate in the index above 1%: Codequiry states that model v3e falsely flags 1.3% of human-written code, with no sample size, and presents it as "low, not zero" (codequiry.com, read 9 November 2026).
Sources
All read 8 November 2026 unless a later date is shown.
- Turnitin AI writing detection FAQs: https://guides.turnitin.com/hc/en-us/articles/28477544839821
- Copyleaks testing methodology: https://copyleaks.com/ai-detector/testing-methodology
- GPTZero FAQ: https://gptzero.me/faq
- Pangram: https://www.pangram.com/
- Winston AI: https://gowinston.ai/
- Originality.ai 5.0 launch post: https://originality.ai/blog/introducing-originality-5-0-and-max
- ZeroGPT DeepAnalyse: https://www.zerogpt.com/deep-analyse
- Proofademic AI text detector: https://proofademic.ai/ai-text-detector/
- Undetectable AI detector: https://undetectable.ai/ai-detector (read 9 November 2026)
- Codequiry AI code detection: https://codequiry.com/detect-ai-written-code (read 9 November 2026)
- Jabarian and Imas 2025, NBER w34223: https://www.nber.org/papers/w34223 (read 9 November 2026)
- Van Vlasselaer et al. 2026: https://doi.org/10.1007/s40979-026-00226-w (read 9 November 2026)
- Dugan et al. 2024, RAID: https://arxiv.org/pdf/2405.07940
- Dugan et al. 2024, RAID, ACL: https://doi.org/10.18653/v1/2024.acl-long.674
- Weber-Wulff et al. 2023: https://doi.org/10.1007/s40979-023-00146-z
- Orenstrakh et al. 2023: https://arxiv.org/abs/2307.07411 (read 9 November 2026)
- Liang et al. 2023: https://doi.org/10.1016/j.patter.2023.100779
- Epoch AI 2026, AI detectors rarely flag human writing: https://epoch.ai/data-insights/ai-detectors-false-negatives (read 9 November 2026)
- GPTZero AI detection benchmarking, 5 February 2026, updated 20 July 2026: https://gptzero.me/news/gptzero-ai-detection-benchmarking-the-industry-standard-in-accuracy-transparency-and-fairness/ (read 9 November 2026)
- Pangram 4.1 model card, 19 October 2026: https://www.pangram.com/research/model-card/pangram-4-1 (read 9 November 2026)
Revision log: 2026-11-09: page created from the vendor claim and published false positive fields of 28 cards at index version 2026.11.4; 2026-11-09: Epoch AI 2026 (0 of 495 for Pangram and GPTZero, 19 of 495 for Originality.ai) and the dated vendor benchmarks (GPTZero 0.08%, Pangram 70 of 781,981) added at index version 2026.11.7; 2026-11-09: published test data and false positive studies questions linked (seventh wave)
Related pages: Is an AI detector itself an AI? The four detection methods behind the cards, each sourced, Is there an AI detector for short texts? Every stated minimum length in the index, AI detector false positive rates, ranked and Questions about AI detectors, answered from the index.