Which AI detectors have the most false positive studies? GPTZero with 11 records, the GPT-2 Output Detector with 8, and every human-text figure the index holds for 15 text cards
GPTZero has the most false positive data of the 72 AI detectors in the index: 11 of its 16 study records print a false positive rate or count on human-written text, counted from the studies list of every card read 9 November 2026 in the data file, running from 9 of 18 human documents (50%) in Weber-Wulff 2023 (DOI 10.1007/s40979-023-00146-z, read 9 November 2026) to 0 of 495 pre-2022 passages in Epoch AI 2026 (epoch.ai, read 9 November 2026). The GPT-2 Output Detector follows with 8 records, Copyleaks and Originality.ai with 6, ZeroGPT with 5, and Turnitin, Pangram and Writer with 4 each. Fifteen text cards have at least one such figure, 4 image cards have one on human-made images, and 23 cards have none, 21 of them because no indexed study tested them at all.
A record counts here when it prints a number for the detector on human-written input: a false positive rate or count, a false accusation count, a specificity, or a true negative rate, each of which names how often human text was called AI. Records that fix the false positive rate by threshold rather than measure it (Krishna 2023 and Sadasivan 2023), records with no human text (Kar 2024 and Chaka 2023), a record whose per-tool rates appear only in a chart (Liang 2023), a record that prints mean scores rather than counts (Alshammari and Rao 2025) and vendor statements (the Turnitin claims record) do not count, and the table says so where it matters. The 2026 false positive evidence review lists every figure with its threshold in one table; the false positive rate ranking picks one figure per card by its stated rule; this page counts how many records stand behind each card. It is one of the questions answered from the index, and the AI detector register carries every card.
| Detector | Records with a human-text false positive figure, of all records | The figures, as each record prints them | Card's published false positive rate (source) |
|---|---|---|---|
| GPTZero | 11 of 16 | Weber-Wulff 2023 9 of 18 (50%); Walters 2023 1 of 42 (2%); Elkhatat 2023 1 of 5 (20%); Perkins 2024 1 of 10 (10%); RAID 2024 0.00% at threshold 0.5; Orenstrakh 2023 52 of 114; Jabarian and Imas 2025 0.0071 of 1,992; Van Vlasselaer 2026 0 of 40; Pegoraro 2023 true negative rate 93.5%; DAMAGE 2025 1.47% at the default threshold; Epoch AI 2026 0 of 495 | 0.0% (RAID 2024) |
| GPT-2 Output Detector | 8 of 12 | Weber-Wulff 2023 2 of 18 (11.1%); Perkins 2024 0 of 10; RAID 2024 6.59% (RoBERTa-Base) and 2.91% (RoBERTa-Large) at 0.5; Orenstrakh 2023 2 of 114; Gao 2023 specificity 94% on 50 abstracts; Jabarian and Imas 2025 0.9503 of 1,992; Pegoraro 2023 true negative rate 96.4%; Ibrahim 2023 16 of 120 human ESL essays below a 95% originality cut, printed under the author's label "false negatives" | 0.0% (Perkins 2024) |
| Copyleaks | 6 of 9 | Walters 2023 0 of 42; Elkhatat 2023 1 of 5 (20%); Perkins 2024 5 of 10 (50%); Orenstrakh 2023 1 of 114, and 10 of 10 Spanish human theses flagged; Van Vlasselaer 2026 0 of 40; Pegoraro 2023 true negative rate 92.1% | 0.0% (Walters 2023) |
| Originality.ai | 6 of 7 | Walters 2023 0 of 42 (2 uncertain); RAID 2024 0.25% at 0.5; Orenstrakh 2023 7 of 114; Jabarian and Imas 2025 0.0011 of 1,992; Pegoraro 2023 true negative rate 95.0%; Epoch AI 2026 19 of 495 (3.84%) | 0.0% (Walters 2023) |
| ZeroGPT | 5 of 7 | Weber-Wulff 2023 0 of 18; Walters 2023 1 of 42 (2%); Perkins 2024 0 of 10; RAID 2024 1.42% at 0.5; Pegoraro 2023 true negative rate 92.2% | 0.0% (Weber-Wulff 2023) |
| Turnitin | 4 of 5 | Weber-Wulff 2023 0 of 18; Walters 2023 0 of 42; Perkins 2024 0 of 10; Van Vlasselaer 2026 0 of 40; the fifth record is the vendor's own statement of under 1% on 700,000 or more papers | 0.0% (Walters 2023) |
| Pangram | 4 of 4 | Jabarian and Imas 2025 0.0000 of 1,992; Van Vlasselaer 2026 0 of 40; DAMAGE 2025 0.27% (LLM Baseline) and 3.40% (DAMAGE) at the default threshold, the vendor's own research models; Epoch AI 2026 0 of 495 | 0.0% (Jabarian and Imas 2025) |
| Writer | 4 of 5 | Weber-Wulff 2023 1 of 18 (5.6%); Walters 2023 14% of 42; Elkhatat 2023 2 of 5 rated Likely AI-Generated and 1 Uncertain; Pegoraro 2023 true negative rate 94.5% | 5.6% (Weber-Wulff 2023) |
| GLTR | 3 of 4 | RAID 2024 99.3% at 0.5 (100% at 0.25, 0.05% at 0.95); Orenstrakh 2023 20 of 114; Pegoraro 2023 true negative rate 98.4% | 99.3% (RAID 2024) |
| BrandWell | 3 of 3 | Weber-Wulff 2023 0 of 18 (tested as Content at Scale); Walters 2023 10% of 42; Pegoraro 2023 true negative rate 79.8% | 0.0% (Weber-Wulff 2023) |
| DetectGPT | 2 of 4 | Weber-Wulff 2023 0 of 18; Pegoraro 2023 true negative rate 80.0% | 0.0% (Weber-Wulff 2023) |
| Winston AI | 2 of 2 | Weber-Wulff 2023 0 of 18; RAID 2024 0.55% at 0.5 | 0.0% (Weber-Wulff 2023) |
| Sapling | 1 of 3 | Walters 2023 0 of 42 (31% uncertain) | 0.0% (Walters 2023) |
| Scribbr | 1 of 1 | Walters 2023 1 of 42 (2%) | 2.0% (Walters 2023) |
| Compilatio | 1 of 1 | Weber-Wulff 2023 0 of 18 | 0.0% (Weber-Wulff 2023) |
Figures as each record prints them, from the studies list and published false positive rate field of each card read 9 November 2026; the record linked in each cell carries the base and threshold.
Which detector has the most false positive measurements?
GPTZero, with 11, and the eleven do not agree. Three records found no false positive at all: RAID 2024 at 0.00% of its human-written documents at threshold 0.5 (DOI 10.18653/v1/2024.acl-long.674, read 9 November 2026), Van Vlasselaer 2026 at 0 of 40 papers of 4,000 words or more (DOI 10.1007/s40979-026-00226-w, read 9 November 2026) and Epoch AI 2026 at 0 of 495 passages of about 500 words (epoch.ai, read 9 November 2026), and Jabarian and Imas 2025 found 0.0071 of 1,992, about 14 passages (NBER w34223, read 9 November 2026). Two found the opposite: Weber-Wulff 2023 recorded 9 false accusations in 18 human documents, 6 of them on machine-translated human text, in March 2023 (DOI 10.1007/s40979-023-00146-z, read 9 November 2026), and Orenstrakh 2023 recorded 52 false positives in 114 human English coursework submissions in April 2023 (arXiv 2307.07411, read 9 November 2026). The false positive rate ranking carries RAID's 0.00% as the card's figure by its stated rule, the lowest study-sourced rate with the larger base breaking ties, and the card itself lists the other ten.
The GPT-2 Output Detector's 8 records split the same way, from 0 of 10 in Perkins 2024 (DOI 10.1186/s41239-024-00487-w, read 9 November 2026) to 0.9503 of 1,992 in Jabarian and Imas 2025, where the 2019 RoBERTa model labelled more than 90% of pre-2020 human passages as AI (NBER w34223, read 9 November 2026). Copyleaks' 6 run from 0 of 42 in Walters 2023 and 0 of 40 in Van Vlasselaer 2026 to 5 of 10 in Perkins 2024 and 10 of 10 Spanish human theses in Orenstrakh 2023 (arXiv 2307.07411, read 9 November 2026). Originality.ai's 6 run from 0.0011 of 1,992 in Jabarian and Imas 2025 to 19 of 495 (3.84%) in Epoch AI 2026, 16 of the 19 fiction (epoch.ai, read 9 November 2026).
Which records measured false positives, and on how much human text?
Thirteen of the 35 records print a human-text figure for at least one card, and two more print one on human-made images. The human bases run from 5 control paragraphs in Elkhatat 2023 (DOI 10.1007/s40979-023-00140-5, read 9 November 2026) and 10 researcher-written samples in Perkins 2024, through 18 documents in Weber-Wulff 2023, 40 papers in Van Vlasselaer 2026, 42 first-year papers in Walters 2023 (DOI 10.1515/opis-2022-0158, read 9 November 2026), 50 journal abstracts in Gao 2023 (DOI 10.1038/s41746-023-00819-6, read 9 November 2026), 114 coursework submissions in Orenstrakh 2023, 120 ESL essays in Ibrahim 2023 (DOI 10.1186/s40468-023-00260-2, read 9 November 2026) and 495 passages in Epoch AI 2026, to 1,992 passages in Jabarian and Imas 2025 and 14,971 human-written documents in RAID 2024. Pegoraro 2023 prints a true negative rate per tool on the human half of a 10% sample of 131,512 responses (arXiv 2304.01487, read 9 November 2026), and DAMAGE 2025 prints a default-threshold rate on human essays without printing the count (ACL Anthology 2025.genaidetect-1.9, read 9 November 2026). On images, Ha and Passananti 2024 scored 280 human artworks (DOI 10.1145/3658644.3670306, read 9 November 2026) and RAID image 2025 scored 50 real images (arXiv 2506.03988, read 9 November 2026). Liang 2023 ran seven detectors on 91 TOEFL essays and reported a 61.22% average false positive rate, but printed per-tool rates only in a chart, so no card carries a number from it (DOI 10.1016/j.patter.2023.100779, read 9 November 2026).
Which detectors have no false positive measurement?
Twenty-three cards. QuillBot and Undetectable AI have study records but no human-text count: Kar 2024 used no human text, and Alshammari and Rao 2025 printed QuillBot's average human score, 100% on 49 human answers, as a mean rather than a count (arXiv 2507.17944, read 9 November 2026). The other 21 have no study record of any kind: Grammarly, Phrasly, JustDone, Proofademic, Walter Writes, AIDetectPlus, Codequiry, WasItAI, Reality Defender, Grammarly Authorship, Turnitin Clarity, SynthID, Smodin, Surfer, Quetext, TruthScan, Deepware, Sensity, Codeleaks, Pindrop Pulse and Resemble Detect, by the study lists on their cards read 9 November 2026. Vendor figures, set out in the question on claims under 1 percent and the view of vendors that publish a false positive rate, are recorded on the cards and do not count as study records here.
Cards with study records outside the text table
| Detector | Study records | Figure on human inputs, as printed | Source |
|---|---|---|---|
| Hive | 2 | Ha and Passananti 2024: 0.00% of 280 human artworks, rising to 3.23% under high-intensity Glaze; RAID image 2025: 1.00 (accuracy) on 50 real images at a 0.5 threshold | DOI 10.1145/3658644.3670306 |
| AI or Not | 1 | Ha and Passananti 2024: 24.47% of 280 human artworks flagged (tested as Optic) | DOI 10.1145/3658644.3670306 |
| Illuminarty | 1 | Ha and Passananti 2024: 67.40% of 280 human artworks flagged | DOI 10.1145/3658644.3670306 |
| Sightengine | 1 | RAID image 2025: 1.00 (accuracy) on 50 real images at a 0.5 threshold | arXiv 2506.03988 |
| QuillBot | 2 | Alshammari and Rao 2025: average human score 100% on 49 human answers, a mean score; Kar 2024: no human text in the sample | arXiv 2507.17944 |
| Undetectable AI | 1 | Kar 2024: no human text in the sample | DOI 10.1177/02537176241247934 |
Figures as each record prints them, read 9 November 2026; image figures are on human-made images, not text.
Eleven records on GPTZero yield eleven different false positive rates, from 0.00% to 50%, because the human bases run from 5 paragraphs to 14,971 documents and the tool changed between March 2023 and June 2026; a count of studies measures attention, not safety.
Questions about false positive studies
Which AI detector has the most false positive studies?
11 of GPTZero's 16 study records print a human-text false positive figure, the most of any card, read 9 November 2026: 0.00% in RAID 2024, 0 of 40 in Van Vlasselaer 2026, 0 of 495 in Epoch AI 2026, 0.0071 of 1,992 in Jabarian and Imas 2025, 1 of 42 in Walters 2023, 1 of 10 in Perkins 2024, 1 of 5 in Elkhatat 2023, 1.47% in DAMAGE 2025, true negative rate 93.5% in Pegoraro 2023, 52 of 114 in Orenstrakh 2023 and 9 of 18 in Weber-Wulff 2023.
How many AI detectors have a measured false positive rate?
15 text cards and 4 image cards of 72, by the study lists read 9 November 2026; the same 15 are the 23 cards whose published false positive rate field comes from a study rather than a vendor. 21 cards have no study record at all.
Which study measured false positives on the most human text?
14,971 human-written documents in RAID 2024, the largest base, printing 0.00% for GPTZero, 0.25% for Originality.ai, 0.55% for Winston AI, 1.42% for ZeroGPT, 2.91% and 6.59% for the two GPT-2 Output Detector models and 99.3% for GLTR at threshold 0.5 (DOI 10.18653/v1/2024.acl-long.674, read 9 November 2026).
Does Turnitin have false positive studies?
4 study records print 0 false positives for Turnitin: 0 of 18 in Weber-Wulff 2023, 0 of 42 in Walters 2023, 0 of 10 in Perkins 2024 and 0 of 40 in Van Vlasselaer 2026, read 9 November 2026; the vendor's own statement of under 1% on 700,000 or more pre-ChatGPT papers is a fifth record and not a study (guides.turnitin.com, read 8 November 2026).
Which detectors have zero false positives in every study?
3 text cards: Turnitin, with 0 of 18, 0 of 42, 0 of 10 and 0 of 40 across 4 records (110 human documents in all), Sapling with 0 of 42 in its 1 record, and Compilatio with 0 of 18 in its 1 record, read 9 November 2026. Winston AI's two records print 0 of 18 and 0.55%, and DetectGPT's print 0 of 18 and a true negative rate of 80.0%.
Sources
All read 9 November 2026 in the primary source unless an earlier date is shown; the record linked in each table cell carries the base and threshold.
- Weber-Wulff 2023: https://doi.org/10.1007/s40979-023-00146-z
- Walters 2023: https://doi.org/10.1515/opis-2022-0158
- Elkhatat 2023: https://doi.org/10.1007/s40979-023-00140-5
- Perkins 2024: https://doi.org/10.1186/s41239-024-00487-w
- RAID 2024 (Dugan et al.): https://doi.org/10.18653/v1/2024.acl-long.674
- Orenstrakh 2023: https://arxiv.org/abs/2307.07411
- Gao 2023: https://doi.org/10.1038/s41746-023-00819-6
- Jabarian and Imas 2025, NBER w34223: https://www.nber.org/papers/w34223
- Van Vlasselaer 2026: https://doi.org/10.1007/s40979-026-00226-w
- Pegoraro 2023: https://arxiv.org/abs/2304.01487
- DAMAGE 2025: https://aclanthology.org/2025.genaidetect-1.9/
- Epoch AI 2026 data insight: https://epoch.ai/data-insights/ai-detectors-false-negatives
- Ibrahim 2023: https://doi.org/10.1186/s40468-023-00260-2
- Liang 2023: https://doi.org/10.1016/j.patter.2023.100779
- Ha and Passananti 2024: https://doi.org/10.1145/3658644.3670306
- RAID image 2025: https://arxiv.org/abs/2506.03988
- Alshammari and Rao 2025: https://arxiv.org/abs/2507.17944
- Kar 2024: https://doi.org/10.1177/02537176241247934
- Turnitin AI writing detection FAQs: https://guides.turnitin.com/hc/en-us/articles/28477544839821-Turnitin-s-AI-writing-detection-capabilities-FAQs (read 8 November 2026)
Revision log: 2026-11-09: page created from the studies lists and published false positive rate fields of 42 cards and 15 study records at index version 2026.11.7.
Related pages: Which AI detectors run in Canvas without Turnitin? The seven cards that name a Canvas integration of their own, how each connects, and what the plan costs, Which AI detectors accept Google Docs links? None take a share link; twelve run inside Google Docs through an add-on or extension, and three of them record the writing history, AI detector false positive rates, ranked and Questions about AI detectors, answered from the index.