AI detectors that publish a false positive rate
A false positive rate is published for 26 of the 72 AI detector cards in the index, by the published_fpr field compiled from the index data file on 9 November 2026: 23 of those figures come from a study in the studies index and 3 come from the vendor's own page, with no indexed study behind them; BrandWell, Compilatio and DetectGPT joined the study rows with the fourth wave of cards on 9 November 2026, each at 0 false accusations in Weber-Wulff 2023's 18 human documents, and GPTKit (0 of 114, Orenstrakh 2023), PlagiarismCheck (0 of 18, Weber-Wulff 2023) and Fast-DetectGPT (23.2%, RAID 2024) joined on 10 November 2026 with the ninth wave, followed later the same day by Binoculars (0.00%, RAID 2024), Check For AI and Writefull (1 of 18 each, Weber-Wulff 2023) with the tenth; RADAR (3.48%, RAID 2024) and SEO.ai (12% of 42, Walters 2023) joined the study rows on 10 November 2026 with the eleventh wave; Ircam Amplify's under-1% music bound joined the vendor table with no study behind it. The figures run from 0 false positives in 42 human-written papers for Copyleaks, Turnitin, Originality.ai and Sapling in Walters 2023 (DOI 10.1515/opis-2022-0158, read 8 November 2026) to 99.3% of RAID's human-written documents for GLTR at a 0.5 threshold (arXiv 2405.07940, read 9 November 2026). The remaining 46 cards publish no rate of either kind; TruthScan's "<1%" is an image figure and stays on its card. This page is one of five slices under AI detector features; the false positive rate ranking sorts the study figures, the annual false positive evidence review lists every figure every study prints, and the home board shows every AI detector in the index on the same fields.
- Field 1Source type
- Field 2False positive figure
- Field 3Base
- Field 4Threshold or setting
- Field 5Study or page, read date
One field, rank_fields.published_fpr, with a source tag. The value is the lowest false positive rate on human-written text that a study in the index prints for the detector; where no indexed study tested the tool and the vendor states a rate, the vendor's figure is recorded with the source "vendor" and never ranked. The table marks each row by that source type, because a 0 of 42 measured by an outside author and a 0.2% measured by the vendor on its own data are different kinds of number (methodology). Vendor statements that sit beside a study figure on the same card, such as Turnitin's "under 1%", appear in the second table so that neither kind is lost.
Which field decides whether a detector publishes a false positive rate?
| Detector | Source type | False positive figure | Base | Threshold or setting | Study or page, read date |
|---|---|---|---|---|---|
 GPTZero | Study | 0.00% | RAID human-written documents | Naive threshold 0.5 (0.03% at 0.25) | RAID 2024, read 8 November 2026 |
 GPT-2 Output Detector | Study | 0 of 10 | 10 human samples written by the researchers | The demo's own verdict | Perkins 2024, read 9 November 2026 |
 Copyleaks | Study | 0 of 42 | 42 human-written first-year papers | The tool's own verdict | Walters 2023, read 8 November 2026 |
 Originality.ai | Study | 0 of 42 | Same 42, 2 returned uncertain | Detection model 1.4, the tool's own verdict | Walters 2023, read 8 November 2026 |
 Turnitin | Study | 0 of 42 | Same 42 | The tool's own verdict | Walters 2023, read 8 November 2026 |
 ZeroGPT | Study | 0 of 18 | 18 human-written documents, 9 machine-translated | The tool's own verdict, March 2023 | Weber-Wulff 2023, read 8 November 2026 |
 Pangram | Study | 0.0000 | 1,992 human passages written before 2020 | 0.5 on the API's prob_ai field (0.0010 at 0.1) | Jabarian and Imas 2025, read 9 November 2026 |
 Sapling | Study | 0 of 42 | Same 42 as Walters 2023, 31% returned uncertain | The tool's own verdict | Walters 2023, read 8 November 2026 |
 Winston AI | Study | 0 of 18 | 18 human-written documents | The tool's own verdict, March 2023 | Weber-Wulff 2023, read 8 November 2026 |
 Scribbr | Study | 2% | 42 human-written first-year papers | The tool's own verdict | Walters 2023, read 8 November 2026 |
 Writer | Study | 1 of 18 (5.6%) | 18 human-written documents, the false accusation on a machine-translated text | The tool's own verdict, March 2023; tool withdrawn 22 December 2025 | Weber-Wulff 2023, read 9 November 2026 |
 GLTR | Study | 99.3% | RAID human-written documents | Naive threshold 0.5, repository defaults (0.05% at 0.95) | RAID 2024, read 9 November 2026 |
 Proofademic | Vendor | 0.2% | 4,200 fully human-authored academic texts, March 2026 evaluation | The reported production threshold | proofademic.ai, read 8 November 2026 |
 Codequiry | Vendor | 1.3% | Human-written code; no sample size published | Model v3e | codequiry.com, read 9 November 2026 |
 Undetectable AI | Vendor | 0% | A study of ten free detectors in the Indian Journal of Psychological Medicine, as the vendor summarises it; no base stated on the page | Not stated | undetectable.ai/ai-detector, read 9 November 2026 |
| DetectGPT | Study | 0 of 18 | 18 human-written documents, 9 machine-translated | The public demo's own verdict, March 2023 | Weber-Wulff 2023, Table 9, read 9 November 2026 |
 BrandWell | Study | 0 of 18 | Same 18, tested as Content at Scale; 36 of 36 AI documents undetected in the same test | The tool's own verdict, March 2023 | Weber-Wulff 2023, Table 9, read 9 November 2026 |
 Compilatio | Study | 0 of 18 | Same 18 | The tool's own verdict, March 2023 | Weber-Wulff 2023, Table 9, read 9 November 2026 |
 GPTKit | Study | 0 of 114 | 114 human-written English computing submissions | The paper's threshold method | Orenstrakh 2023, Table 8, read 9 November 2026 |
 PlagiarismCheck | Study | 0 of 18 | Same 18 human documents as the other Weber-Wulff rows | The tool's own verdict, May 2023 | Weber-Wulff 2023, Table 9, read 9 November 2026 |
| Fast-DetectGPT | Study | 23.2% | RAID human-written documents | Naive threshold 0.5 (47.3% at 0.25, 1.70% at 0.95), repository defaults | RAID 2024, Table 4, read 10 November 2026 |
 Binoculars | Study | 0.00% | RAID human-written documents | Naive threshold 0.5 (0.07% at 0.25), repository defaults with the Falcon-7B pair | RAID 2024, Table 4, read 10 November 2026 |
 Check For AI | Study | 5.6% | 1 of 18 human-written documents, 9 of them machine-translated | The tool's own verdict, March 2023 | Weber-Wulff 2023, Table 9, read 10 November 2026 |
 Writefull | Study | 5.6% | 1 of 18 human-written documents, 9 of them machine-translated | The tool's own verdict, March 2023 | Weber-Wulff 2023, Table 9, read 10 November 2026 |
 RADAR | Study | 3.48% | RAID human-written documents | Naive threshold 0.5 (7.48% at 0.25, 1.23% at 0.95), queried from Hugging Face | RAID 2024, Table 4, read 10 November 2026 |
 SEO.ai | Study | 12% | 42 human-written papers | The tool's own verdict, June to July 2023 | Walters 2023, Table 4, read 8 November 2026 |
Values as each study or vendor page states them; the study figures are checked against the index data file at every build. A "Study" row is the lowest figure an indexed study prints for that tool on human-written text; a "Vendor" row is the vendor's own figure on data the vendor chose. AI or Not's one measured rate, 24.47% of 280 human artworks in Ha and Passananti 2024, is an image result and is carried on the AI or Not card and in the false negative rate ranking rather than in this field, which holds rates on human-written text.
Vendor statements beside a study figure
Why each card is in this view
- AIDI-D-0001GPTZero
0.00% of RAID's human-written documents at a 0.5 threshold (arXiv 2405.07940, Table 4, read 8 November 2026); the same card carries 9 of 18 in Weber-Wulff 2023 and 52 of 114 in Orenstrakh 2023, listed on the ranking.
- AIDI-D-0019GPT-2 Output Detector
0 false accusations in 10 human samples (arXiv 2403.19148, Table 10, read 9 November 2026); 0.9503 of 1,992 passages for the RoBERTa-base weights in Jabarian and Imas 2025 sits beside it.
- AIDI-D-0003Copyleaks
0 of 42 human papers (De Gruyter 10.1515/opis-2022-0158, Table 4, read 8 November 2026); the methodology page's 0.30% is a vendor figure.
- AIDI-D-0004Originality.ai
0 of 42 with 2 uncertain, detection model 1.4 (De Gruyter PDF, read 8 November 2026).
- AIDI-D-0002Turnitin
0 of 42 (De Gruyter PDF, read 8 November 2026), with the author's note that the papers may have been in the training data.
- AIDI-D-0011ZeroGPT
0 of 18 (arXiv 2306.15666, Table 9, read 8 November 2026); RAID 2024 recorded 1.42% at 0.5.
- AIDI-D-0006Pangram
0.0000 of 1,992 at a 0.5 threshold (NBER w34223, Table 3, read 9 November 2026); the vendor's 1 in 10,000 is in the second table.
- AIDI-D-0010Sapling
0 of 42 with 31% returned uncertain (De Gruyter PDF, read 8 November 2026).
- AIDI-D-0005Winston AI
0 of 18 (arXiv 2306.15666, Table 9, read 8 November 2026); the vendor's zero across 27,743 documents is in the second table.
- AIDI-D-0009Scribbr
2% of 42 papers, one paper (De Gruyter PDF, read 8 November 2026).
- AIDI-D-0023Writer
1 of 18, on a machine-translated text (arXiv 2306.15666, Table 9, read 9 November 2026); Walters 2023 recorded 14% of 42 and the tool was withdrawn on 22 December 2025.
- AIDI-D-0017GLTR
99.3% at the default threshold of 0.5, 0.05% at 0.95 (arXiv 2405.07940, Table 4, read 9 November 2026).
- AIDI-D-0014Proofademic
the vendor states 0.2% on 4,200 human-authored academic texts in a March 2026 evaluation (proofademic.ai, read 8 November 2026); no indexed study tested the tool.
- AIDI-D-0020Codequiry
the vendor states 1.3% of human-written code falsely flagged by model v3e with no sample size (codequiry.com, read 9 November 2026); no indexed study tested the tool.
- AIDI-D-0024Undetectable AI
the vendor's comparison table states 0% false positives, citing a study of ten free detectors that the index has not yet read (undetectable.ai, read 9 November 2026); no indexed study tested the tool.
- AIDI-D-0035DetectGPT
0 false accusations in 18 human documents through the public demo (arXiv 2306.15666, Table 9, read 9 November 2026); Krishna 2023 and Sadasivan 2023 fixed the false positive rate at 1% by threshold rather than measuring it.
- AIDI-D-0030BrandWell
0 of 18 as Content at Scale, beside 36 of 36 AI documents undetected in the same test (arXiv 2306.15666, Tables 9 and 10, read 9 November 2026); Walters 2023 recorded 10% of 42 human papers called AI for the same tool.
- AIDI-D-0032Compilatio
0 of 18 in Weber-Wulff 2023 (arXiv 2306.15666, Table 9, read 9 November 2026); the vendor's own below 1% on 3,700 human texts sits in the vendor table (support.compilatio.net, read 9 November 2026).
- AIDI-D-0045GPTKit
0 of 114, the only detector of eight at zero in Orenstrakh 2023 (arXiv 2307.07411, Table 8, read 9 November 2026); 0 of 10 in Perkins 2024 on the same card.
- AIDI-D-0046PlagiarismCheck
0 of 18 in Weber-Wulff 2023 (arXiv 2306.15666, Table 9, read 9 November 2026), with the vendor's own sentence-level and document-level rates in the second table because they sit beside a study figure.
- AIDI-D-0049Fast-DetectGPT
23.2% at the naive 0.5 threshold in RAID 2024 (arXiv 2405.07940, Table 4, read 10 November 2026), the GLTR treatment of a repository-default run. Isgen's "nearly 0%" and Detecting-AI.com's "under 1%" are bounds with no base and enter neither table (the TruthScan treatment).
- AIDI-D-0053Binoculars
0.00% of RAID's human-written documents at threshold 0.5 and 0.07% at 0.25, run from the official repository (arXiv 2405.07940, Table 4, read 10 November 2026); DAMAGE 2025's 5.40% at the default threshold has no printed essay count and stays in the card's base.
- AIDI-D-0058Check For AI
1 false accusation in 18 human documents in Weber-Wulff 2023 (arXiv 2306.15666, Table 9, read 10 November 2026); the service was sunset in July 2023.
- AIDI-D-0059Writefull
1 false accusation in 18 human documents in Weber-Wulff 2023 (Table 9), with Pegoraro 2023's 99.3% true negative rate quoted and not inverted (arXiv 2304.01487, Table I, read 10 November 2026); the tool is discontinued.
- AIDI-D-0071RADAR
3.48% of RAID's human-written documents at threshold 0.5, queried from Hugging Face as TrustSafeAI/RADAR-Vicuna-7B (arXiv 2405.07940, Table 4, read 10 November 2026); DAMAGE 2025's 51.87% at the default threshold stays in the base because the essay count is not printed.
- AIDI-D-0069SEO.ai
12% of 42 human-written papers called AI in Walters 2023, the third-highest false positive rate of the 16 detectors in that study (DOI 10.1515/opis-2022-0158, Table 4, read 8 November 2026). GPT Radar, Grammica and IvyPanda, added the same day from the same record, carry no rate because the record names a human-paper figure only for the eight detectors it discusses, and the GPT-2 Output Detector's base gained Rashidi 2023's up to 8.69% of 1,600 real abstracts while its field keeps the lowest study value.
What the evidence says about published false positive rates
The false positive studies question, added 9 November 2026, counts the records behind each study-sourced figure in this view and tables the image rows for AI or Not and Illuminarty apart.
Fifteen cards publish a rate and the rates do not travel. GPTZero's row reads 0.00% because RAID 2024 is the largest human sample in the index, at 14,971 documents (DOI 10.18653/v1/2024.acl-long.674, read 8 November 2026), and the same card carries 9 of 18 in Weber-Wulff 2023 (DOI 10.1007/s40979-023-00146-z, read 8 November 2026) and 52 of 114 in Orenstrakh 2023 (arXiv 2307.07411, read 9 November 2026). The three vendor rows share one property: each rests on data the vendor chose, and Undetectable AI's 0% names a journal study without stating its base, so the index files it as a statement until the paper is indexed.
The vendor statements in the second table are not false; they are unverified by anyone outside the company, on this record. Turnitin's "under 1%" on 700,000 pre-ChatGPT papers is the largest base any vendor names, and the four independent results for Turnitin (0 of 42, 0 of 18, 0 of 10 and 0 of 40) are consistent with it (guides.turnitin.com, read 8 November 2026; Van Vlasselaer 2026). Pangram's 1 in 10,000 and its study figure of 0.0000 of 1,992 point the same way (NBER w34223, read 9 November 2026). GPTZero's 1% on RAID and the 0.00% that RAID itself prints at 0.5 differ because the vendor quotes the rate at a fixed 1% operating point (gptzero.me/faq, read 8 November 2026).
A published rate is one number from one sample on one date, and the thirteen cards that publish none are not worse detectors for it, only less documented. No detector score should be the only evidence in a misconduct, hiring or publishing decision.
Sources
- https://arxiv.org/pdf/2405.07940 (read 8 November 2026)
- https://doi.org/10.18653/v1/2024.acl-long.674 (read 8 November 2026)
- https://www.degruyterbrill.com/document/doi/10.1515/opis-2022-0158/pdf (read 8 November 2026)
- https://doi.org/10.1515/opis-2022-0158 (read 8 November 2026)
- https://arxiv.org/pdf/2306.15666 (read 8 November 2026)
- https://doi.org/10.1007/s40979-023-00146-z (read 8 November 2026)
- https://arxiv.org/pdf/2403.19148 (read 8 November 2026)
- https://www.nber.org/papers/w34223 (read 9 November 2026)
- https://arxiv.org/abs/2307.07411 (read 9 November 2026)
- https://arxiv.org/abs/2402.03214 (read 9 November 2026)
- https://guides.turnitin.com/hc/en-us/articles/28477544839821 (read 8 November 2026)
- https://copyleaks.com/ai-detector/testing-methodology (read 8 November 2026)
- https://gptzero.me/faq (read 8 November 2026)
- https://www.pangram.com/ (read 8 November 2026)
- https://gowinston.ai/ (read 8 November 2026)
- https://originality.ai/blog/introducing-originality-5-0-and-max (read 8 November 2026)
- https://www.zerogpt.com/deep-analyse (read 8 November 2026)
- https://proofademic.ai/ai-text-detector/ (read 8 November 2026)
- https://codequiry.com/detect-ai-written-code (read 9 November 2026)
- https://undetectable.ai/ai-detector (read 9 November 2026)
Revision log: 2026-11-09: view created, 15 of 28 cards; 2026-11-09: DetectGPT, BrandWell and Compilatio added from Weber-Wulff 2023, Compilatio's vendor rate to the second table, 18 of 38 cards2026-11-09: false positive studies question linked (seventh wave)2026-11-10: GPTKit, PlagiarismCheck and Fast-DetectGPT added to the study rows and PlagiarismCheck's vendor rates to the second table (ninth wave)2026-11-10: Binoculars, Check For AI and Writefull added to the study rows and Ircam Amplify's vendor rates to the second table (tenth wave)2026-11-10: RADAR and SEO.ai added to the study rows (eleventh wave); Rashidi 2023 added to the GPT-2 Output Detector base
FAQ: published false positive rates
How many AI detectors publish a false positive rate?
26 of the 72 cards in the index as of 9 November 2026: 23 with a figure from an indexed study and 3 (Proofademic, Codequiry and Undetectable AI) with a vendor figure only. The other 13 publish no rate.
Which AI detector vendors publish their own false positive rate?
Ten vendors state a figure on their own pages, read 8 and 9 November 2026: Turnitin (under 1%), Copyleaks (0.30%), GPTZero (1%, and 1 in 10,402), Pangram (1 in 10,000), Winston AI (zero across 27,743 documents), Originality.ai (99.954% specificity), ZeroGPT (under 1%), Proofademic (0.2%), Codequiry (1.3%) and Undetectable AI (0%). The first seven also have a study figure on the card; the last three do not.
Which published false positive rate has the largest base?
RAID 2024's 14,971 human-written documents behind GPTZero's 0.00%, Originality.ai's 0.25%, Winston AI's 0.55%, ZeroGPT's 1.42% and GLTR's 99.3% at threshold 0.5 (arXiv 2405.07940, read 8 November 2026). Among vendor statements, Turnitin's 700,000 pre-ChatGPT papers is the largest named base (guides.turnitin.com, read 8 November 2026).
Why is AI or Not not in this table when a study measured its false positives?
Because the measured rate, 24.47% of 280 human artworks in Ha and Passananti 2024 (arXiv 2402.03214, read 9 November 2026), is an image result and this field holds rates on human-written text, the rule that also keeps Hive's 0% on the same artworks out. Both figures sit on the cards and in the false negative rate ranking.
Which cards publish no false positive rate at all?
Thirteen as of 9 November 2026: QuillBot, Grammarly, Phrasly, JustDone, Walter Writes, Hive, AIDetectPlus, WasItAI, Reality Defender, AI or Not, Grammarly Authorship, Turnitin Clarity and SynthID. Grammarly Authorship and Turnitin Clarity record a writing process and return no score, and SynthID checks for a watermark, so none of the three has a rate to publish.
