2026 review: every published AI detector false positive figure in the index
Twelve of the 72 detectors in the index have a false positive figure printed by at least one of the 35 indexed studies, by the count compiled from the index data file on 9 November 2026, and the figures run from 0 of 40 human papers for GPTZero, Pangram, Copyleaks and Turnitin in Van Vlasselaer 2026 (DOI 10.1007/s40979-026-00226-w, read 9 November 2026) to 0.9503 of 1,992 human passages for the RoBERTa-base baseline in Jabarian and Imas 2025 (NBER w34223, read 9 November 2026). This review lists every one of those figures in a single table, then records what changed in the index between launch on 8 November 2026 and this review.
Every figure below was checked against the index data file on 9 November 2026. A false positive is a human-written text labelled AI. The table takes each figure as its study prints it, a percentage or a count out of a base, at the threshold the study used, so a 0 of 5 and a 0.0000 of 1,992 sit in the same column with their bases visible. Rows run study by study in the order the records were indexed; the false positive rate ranking sorts the lowest figure per tool, and the false negative rate ranking shows the miss rate from the same studies.
| Study and date | Detector | False positive figure | Base and threshold | Record |
|---|---|---|---|---|
| Weber-Wulff 2023, March 2023 | GPTZero | 9 of 18 (50%) | 18 human-written documents, 9 of them machine-translated; the tool's own verdict | Weber-Wulff 2023 record |
| Weber-Wulff 2023 | GPT-2 Output Detector | 2 of 18 (11.1%) | Same 18; both on machine-translated texts | Weber-Wulff 2023 record |
| Weber-Wulff 2023 | ZeroGPT, Turnitin, Winston AI | 0 of 18 | Same 18 | Weber-Wulff 2023 record |
| Liang 2023, March 2023 | GPTZero, ZeroGPT, Originality.ai, Sapling, GPT-2 Output Detector and two tools without a card | 61.22% average across seven tools | 91 TOEFL essays; 5.19% on 88 US eighth-grade essays; per-tool rates in a chart only | Liang 2023 record |
| Walters 2023, June to July 2023 | GPTZero, Scribbr, ZeroGPT | 2% (1 of 42) | 42 human-written first-year papers; the tool's own verdict, "uncertain" allowed | Walters 2023 record |
| Walters 2023 | Copyleaks, Turnitin, Originality.ai, Sapling | 0 of 42 | Same 42; Originality.ai 2 uncertain, Sapling 31% uncertain | Walters 2023 record |
| Elkhatat 2023 | GPTZero, Copyleaks | 1 of 5 (20%) | 5 human control paragraphs | Elkhatat 2023 record |
| Perkins 2024 | Copyleaks | 5 of 10 (50%) | 10 human samples written by the researchers | Perkins 2024 record |
| Perkins 2024 | GPTZero | 1 of 10 (10%) | Same 10 | Perkins 2024 record |
| Perkins 2024 | ZeroGPT, Turnitin, GPT-2 Output Detector | 0 of 10 | Same 10 | Perkins 2024 record |
| RAID 2024 | GPTZero | 0.03%, 0.00%, 0.00%, 0.00% | RAID human-written documents at thresholds 0.25, 0.5, 0.75, 0.95 | RAID 2024 record |
| RAID 2024 | Originality.ai | 0.47%, 0.25%, 0.17%, 0.07% | Same thresholds | RAID 2024 record |
| RAID 2024 | Winston AI | 0.75%, 0.55%, 0.38%, 0.21% | Same thresholds | RAID 2024 record |
| RAID 2024 | ZeroGPT | 1.71%, 1.42%, 1.21%, 0.90% | Same thresholds | RAID 2024 record |
| RAID 2024 | GPT-2 Output Detector | 6.59% (RoBERTa-base), 2.91% (RoBERTa-Large) | Threshold 0.5 | RAID 2024 record |
| RAID 2024 | GLTR | 100%, 99.3%, 21.0%, 0.05% | Same thresholds, repository default settings | RAID 2024 record |
| Orenstrakh 2023, April 2023 | GPTZero | 52 of 114 | 114 human English coursework submissions; 50% on the authors' probability scale | Orenstrakh 2023 record |
| Orenstrakh 2023 | GLTR | 20 of 114 | Same 114; the authors' z-score wrapper | Orenstrakh 2023 record |
| Orenstrakh 2023 | Originality.ai | 7 of 114 | Same 114 | Orenstrakh 2023 record |
| Orenstrakh 2023 | GPT-2 Output Detector | 2 of 114 | Same 114 | Orenstrakh 2023 record |
| Orenstrakh 2023 | Copyleaks | 1 of 114; 10 of 10 Spanish papers flagged | Same 114; 10 Spanish human theses | Orenstrakh 2023 record |
| Gao 2023, December 2022 | GPT-2 Output Detector | Specificity 94% | 50 journal abstracts; 1.46% fake-score cutoff | Gao 2023 record |
| Jabarian and Imas 2025 | Pangram | 0.0000 | 1,992 human passages written before 2020; threshold 0.5 (0.0010 at 0.1) | Jabarian and Imas 2025 record |
| Jabarian and Imas 2025 | Originality.ai | 0.0011 | Same 1,992; threshold 0.5 (0.0027 at 0.1) | Jabarian and Imas 2025 record |
| Jabarian and Imas 2025 | GPTZero | 0.0071 | Same 1,992; every threshold from 0.1 to 0.9 | Jabarian and Imas 2025 record |
| Jabarian and Imas 2025 | GPT-2 Output Detector (RoBERTa-base baseline) | 0.9503 | Same 1,992; threshold 0.5 (0.8991 at 0.9) | Jabarian and Imas 2025 record |
| Van Vlasselaer 2026, May 2025 | GPTZero, Pangram, Copyleaks, Turnitin | 0 of 40 | 40 human papers of 4,000 words or more; a score under 20% counted as human | Van Vlasselaer 2026 record |
| Ha and Passananti 2024 | Hive | 0.00%; 3.23% under high Glaze | 280 human artworks; a score under 40% counted as human (image detection) | Ha and Passananti 2024 record |
| Epoch AI 2026, June to July 2026 | Pangram, GPTZero | 0 of 495 (0.00%) | 495 passages of about 500 words by 99 authors, all published before 30 November 2022; each detector's default document verdict, Mixed would count as a flag (none occurred); 90% CI 0.00% to 0.54% | Epoch AI 2026 record |
| Epoch AI 2026 | Originality.ai | 19 of 495 (3.84%) | Same 495; 16 of the 19 were fiction, 2 blog, 1 scientific; Turbo 3.0.2 through the web bulk uploader | Epoch AI 2026 record |
| Ibrahim 2023, 2023 | GPT-2 Output Detector | 16 of 120 (13.3%) | 120 human ESL research essays written before ChatGPT, scored below a 95% originality cut set by the author, who labels these "false negatives"; the figure is not entered as a rate on the card | Ibrahim 2023 record |
| DAMAGE 2025, 2025 | GPTZero | 1.47% at the default threshold | Human student essays, count not printed; the paper is authored by Pangram Labs, and the figure does not enter the card's false positive field for want of a base | DAMAGE 2025 record |
| Pegoraro 2023, March 2023 | GPTZero, Copyleaks, Originality.ai, ZeroGPT, GLTR, GPT-2 Output Detector, Writer, BrandWell (as Content at Scale), DetectGPT | True negative rate as printed: 93.5%, 92.1%, 95.0%, 92.2%, 98.4%, 96.4%, 94.5%, 79.8%, 80.0% | The human half of a 10% sample of 131,512 question-and-answer responses (58,546 human in the full benchmark); each tool's own verdict; the paper prints the true negative rate and not a false positive rate, so the figures are quoted as printed and not inverted | Pegoraro 2023 record |
| Ha and Passananti 2024 | AI or Not (as Optic) | 24.47% (image) | 280 human artworks; a score under 40% counted as human (image detection) | Ha and Passananti 2024 record |
| Ha and Passananti 2024 | Illuminarty | 67.40% (image) | Same 280 human artworks, the highest false positive rate of the three commercial image detectors (image detection) | Ha and Passananti 2024 record |
The DAMAGE 2025 record prints a 1.47% default-threshold false positive rate for GPTZero on human student essays without the essay count, tabled above since 9 November 2026 with that note and kept out of the card's false positive field for want of a base; Pegoraro 2023's true negative rates, and the Ha and Passananti 2024 image rows for AI or Not and Illuminarty, joined the table the same day and are labelled as printed; Kar 2024 and Alshammari and Rao 2025 scored no human text as a counted set; the RAID image 2025, ASVspoof 5, SAFE 2025 and DFDC 2020 records score images, speech and video, not written text; the Vanderbilt 2023 statement and the GPTZero and Pangram vendor benchmarks are not tests by an outside party (Vanderbilt statement, vendor benchmarks). Sadasivan 2023 and Krishna 2023 fixed the false positive rate at 1% by threshold and report detection at that rate, so they print no false positive figure for a product; Chaka 2023 ran AI text only; the NIST pilot names no product. Vendor statements sit on the cards and in the false positive rate ranking's vendor table: Turnitin under 1% on over 700,000 papers, Copyleaks 0.30% on 326,120 texts (923 flagged) on its methodology page and .03% on its product page, GPTZero 1% on RAID and 1 in 10,402 in its October 2026 post, Pangram 1 in 10,000, Proofademic 0.2% on 4,200 texts, Codequiry 1.3% with no sample, all read 8 and 9 November 2026 and none of them ranked.
What does the whole table say?
Three things hold across every study. First, the same tool posts a zero and a high figure depending on the text: GPTZero recorded 0 of 40 in Van Vlasselaer 2026 and 0.00% at threshold 0.5 in RAID 2024, and 9 of 18 in Weber-Wulff 2023 and 52 of 114 in Orenstrakh 2023, with the high figures falling on machine-translated texts and on student coursework. Second, the base is small in most 2023 papers, 5, 10, 18 or 42 texts, and large only in RAID 2024 and Jabarian and Imas 2025, so a 0 of 10 cannot rule out a rate of several percent. Third, non-native and translated writing drives the worst figures: Liang 2023's 61.22% average on TOEFL essays (DOI 10.1016/j.patter.2023.100779, read 8 November 2026), the 6 machine-translated texts among GPTZero's 9 false accusations in Weber-Wulff 2023 (DOI 10.1007/s40979-023-00146-z, read 8 November 2026), and Copyleaks flagging all 10 Spanish human theses in Orenstrakh 2023 (arXiv 2307.07411, read 9 November 2026). The per-card count of false positive studies, added 9 November 2026, counts how many of these records stand behind each card: GPTZero 11 of its 16 records, the GPT-2 Output Detector 8 of 12, Copyleaks 6 of 9.
What changed in the index since launch?
The index launched on 8 November 2026 with 16 cards and 8 study records. On 9 November 2026 six cards were added (GLTR, AIDetectPlus, the GPT-2 Output Detector, Codequiry, WasItAI and Reality Defender) and the Chaka 2023 record followed, then eight more records the same day: Sadasivan 2023, Krishna 2023, Orenstrakh 2023, Gao 2023, Jabarian and Imas 2025, Van Vlasselaer 2026, NIST AI 700-1 and Ha and Passananti 2024, bringing the index to 22 cards and 17 records at version 2026.11.3. Six more cards followed later on 9 November 2026 at version 2026.11.4 (Writer, Undetectable AI, AI or Not, Grammarly Authorship, Turnitin Clarity and SynthID), for 72 cards and 35 records at version 2026.11.13; Writer brought a study-sourced rate of 1 in 18 from Weber-Wulff 2023 and 14% of 42 from Walters 2023, listed on the published false positive rate view, and the Q4 2026 index update and studies digest record the quarter.
Those eight records changed four things in the false positive evidence. Pangram gained a study-sourced rate, 0.0000 of 1,992 at threshold 0.5, and moved from the vendor table into the false positive rate ranking, which now ranks 23 tools. Orenstrakh 2023 added the largest 2023 base, 114 human submissions, and with it GPTZero's 52 false positives and Copyleaks' 10 of 10 Spanish papers. Jabarian and Imas 2025 added the first fixed-threshold rates since RAID and the first figure for the GPT-2 Output Detector's RoBERTa-base weights on modern text, 0.9503. Van Vlasselaer 2026 added the first 2026 result, 0 of 40 for all four tools tested. Study counts rose with the records: GPTZero from 7 to 11, the GPT-2 Output Detector from 4 to 9, Copyleaks from 4 to 6, Turnitin and Originality.ai to 5, GLTR to 3, and Pangram and Hive from none to 2 and 1, so the order of the ranking by independent testing and the tie-breaks of the false positive ranking were re-sorted on 9 November 2026 and the change written to each revision log.
Two new rankings read the same records from the other side. The false negative rate ranking shows the miss rate beside every false positive figure above, and the robustness ranking shows how the figures moved under paraphrase. No vendor claim changed between 8 and 9 November 2026; the next recheck of every card is 8 December 2026.
The sixth wave of 9 November 2026 added eleven records and two rows to the table. Epoch AI 2026 is the first record to print zero false positives for two detectors on a base of 495 human passages, for Pangram 3.3.2 and GPTZero model 2026-05-11-base, with Originality.ai Turbo 3.0.2 at 19 of 495 (epoch.ai, read 9 November 2026); under the ranking's lowest-rate rule none of the three replaced the rate already on the card, because a 0 on 495 is not lower than a 0 on 1,992 or on RAID's human set, and the figures were appended to each card's base text instead. Ibrahim 2023 prints 16 of 120 human ESL essays under a 95% originality cut as "false negatives", the author's inverted label; the row above records the count and the index enters no rate from it (DOI 10.1186/s40468-023-00260-2, read 9 November 2026). The vendor figures GPTZero and Pangram publish for themselves, 0.08% and 70 of 781,981, are recorded as dated vendor measurements in a separate record and never join this table (vendor benchmarks).
The second annual piece is the 2027 census, which counts the cards, the records and the held rows on 10 November 2026; the Q3 2027 studies digest lists the five records added since the Q2 digest.
Questions about the false positive evidence
Which AI detectors have published false positive data from a study?
Twelve as of 9 November 2026: GPTZero, the GPT-2 Output Detector, Copyleaks, Originality.ai, Turnitin, ZeroGPT, Pangram, Sapling, Winston AI, Scribbr, GLTR and Hive, each with the figure, base and threshold in the table above and on its card. The other ten cards have a vendor statement or nothing.
Which study has the largest human sample?
RAID 2024 reports rates on 14,971 human-written documents (DOI 10.18653/v1/2024.acl-long.674, read 8 November 2026); Jabarian and Imas 2025 used 1,992 human passages (NBER w34223, read 9 November 2026); Orenstrakh 2023 used 114. Every other study used 42 or fewer.
What will the 2027 review add?
Every record indexed between this review and the next, with the same table, and any card whose vendor statement changes at a monthly recheck, the next on 8 December 2026. Accepted additions are logged in the corrections log with their date.
Sources
All read 8 November 2026 unless a later date is shown.
- Weber-Wulff et al. 2023: https://doi.org/10.1007/s40979-023-00146-z
- Liang et al. 2023: https://doi.org/10.1016/j.patter.2023.100779
- Walters 2023: https://doi.org/10.1515/opis-2022-0158
- Elkhatat et al. 2023: https://doi.org/10.1007/s40979-023-00140-5
- Perkins et al. 2024: https://doi.org/10.1186/s41239-024-00487-w
- Dugan et al. 2024 (RAID): https://doi.org/10.18653/v1/2024.acl-long.674
- Orenstrakh et al. 2023: https://arxiv.org/abs/2307.07411 (read 9 November 2026)
- Gao et al. 2023: https://doi.org/10.1038/s41746-023-00819-6 (read 9 November 2026)
- Jabarian and Imas 2025: 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)
- Ha, Passananti et al. 2024: https://arxiv.org/abs/2402.03214 (read 9 November 2026)
- Sadasivan et al. 2023: https://arxiv.org/abs/2303.11156 (read 9 November 2026)
- Krishna et al. 2023: https://arxiv.org/abs/2303.13408 (read 9 November 2026)
- Turnitin AI writing detection FAQs: https://guides.turnitin.com/hc/en-us/articles/28477544839821
- Copyleaks testing methodology: https://copyleaks.com/ai-detector/testing-methodology
- Epoch AI 2026: https://epoch.ai/data-insights/ai-detectors-false-negatives (read 9 November 2026)
- Ibrahim 2023: https://doi.org/10.1186/s40468-023-00260-2 (read 9 November 2026)
Revision log: 2026-11-09: review created at index version 2026.11.3; 2026-11-09: Epoch AI 2026 and Ibrahim 2023 rows added and the sixth-wave paragraph written at index version 2026.11.7; 2026-11-09: DAMAGE 2025 (GPTZero 1.47%, human count not printed), Pegoraro 2023 true negative rates as printed, and the Ha and Passananti 2024 image rows for AI or Not (24.47%) and Illuminarty (67.40%) added to the table; the per-card false positive studies question linked (seventh wave); 2026-11-10: twelfth-wave pages linked
Related pages: Index update, Q4 2026: twelve cards added, five fields added, four corrections filed, Studies digest, Q4 2026: the nine evaluations added since launch, with one headline figure each, AI detector false positive rates, ranked, How accurate are AI detectors? Every study, indexed and Index updates: dated reviews of the AI detector data.