AI detectors ranked by independent testing Two records added in the fifth wave on 9 November 2026, Pegoraro 2023 and Fleckenstein 2024, changed nine counts and none: Pegoraro 2023 tested ZeroGPT, GPTZero, Copyleaks, Originality.ai, Writer, Content at Scale (BrandWell), DetectGPT, the GPT-2 Output Detector and GLTR, lifting GPTZero to 12, the GPT-2 Output Detector to 10, Copyleaks to 7 and Originality.ai and ZeroGPT to 6; Fleckenstein 2024 tested 289 teachers and no software, so it counts for no card ([Fleckenstein 2024 record](/studies/fleckenstein-2024-teachers/)). Illuminarty, one of four cards added in the same wave, joined with its one image record from Ha and Passananti 2024, so 18 of 72 cards now qualify.
GPTZero appears in 12 of the 35 study records in the index, more than any other AI detector, by the count compiled on 9 November 2026 from the papers linked on this page. 35 of the 72 detectors have at least one study record; the other 37 have none. Two cards added on 9 November 2026, the GPT-2 Output Detector with 4 records and GLTR with 1, joined the ranking on that date; the Chaka 2023 record, added the same day, raised GPTZero to 7, Copyleaks to 4 and GLTR to 2. Eight records added later on 9 November 2026 (Sadasivan 2023, Krishna 2023, Orenstrakh 2023, Gao 2023, Jabarian and Imas 2025, Van Vlasselaer 2026, NIST AI 700-1 and Ha and Passananti 2024) raised GPTZero to 11, the GPT-2 Output Detector to 9, Copyleaks to 6, Turnitin and Originality.ai to 5 and GLTR to 3, and brought Pangram (2) and Hive (1) into the ranking. Six cards added on 9 November 2026 in the third wave brought two more tools in: Writer, whose withdrawn detector was tested in 4 records (Weber-Wulff 2023, Walters 2023, Elkhatat 2023 and Chaka 2023), and AI or Not, tested as Optic in Ha and Passananti 2024; the other four (Undetectable AI, Grammarly Authorship, Turnitin Clarity and SynthID) have no record. Ten cards added in the fourth wave the same day brought three more in: DetectGPT with 3 records (Weber-Wulff 2023, Sadasivan 2023 and Krishna 2023), BrandWell with 2 (Weber-Wulff 2023 and Walters 2023, both testing the tool as Content at Scale) and Compilatio with 1 (Weber-Wulff 2023); the other seven (Smodin, Surfer, Quetext, TruthScan, Sightengine, Deepware and Sensity) have none.
Updated Studies dated 2023 to 2026
Short answer
GPTZero has 16 study records, the GPT-2 Output Detector has 12, Copyleaks has 9, Originality.ai and ZeroGPT have 7 each, Turnitin and Writer have 5, Pangram, GLTR and DetectGPT have 4, Sapling and BrandWell have 3, and no other detector in the index has more than 2; eleven records added in the sixth wave on 9 November 2026 lifted eight of those counts and brought QuillBot, Undetectable AI and Sightengine into the table (see the lead), and before that nine counts rose by one on 9 November 2026 when Pegoraro 2023, a 20-method sweep on ChatGPT and human question-and-answer text, was added (arXiv 2304.01487, read 9 November 2026). 35 of 72 detectors qualify as of 9 November 2026. A high count means a tool has been examined often, not that the tool scored well: GPTZero's records run from 9 false accusations in 18 human texts (Weber-Wulff 2023) to a 0.00% false positive rate at threshold 0.5 (RAID 2024). Sources read 8 November 2026. No detector score should be the only evidence in a misconduct or hiring decision.
). Illuminarty, one of four cards added in the same wave, joined with its one image record from Ha and Passananti 2024, so 18 of 72 cards now qualify.: register plate with the page title and its register ID](/img/independently-tested-1600.webp)
Which AI detectors have been independently tested?
| Rank | Detector | Study records | Most recent study | Best for | Main limitation |
|---|---|---|---|---|---|
| 1 |  GPTZero | 16 | 2026 | Readers weighing many outside results | 9 false accusations in 18 human documents, Weber-Wulff 2023; 52 of 114 in Orenstrakh 2023 |
| 2 |  GPT-2 Output Detector | 13 | 2025 | Researchers who want a free, open baseline | Built for GPT-2; false positive rate 0.9503 of 1,992 passages, Jabarian and Imas 2025 |
| 3 |  Copyleaks | 9 | 2026 | Schools that need LMS integrations | 5 false accusations in 10 human samples, Perkins 2024; 75% of 40 fully AI papers missed, Van Vlasselaer 2026 |
| 4 |  Originality.ai | 7 | 2026 | Publishers checking long-form text | 9.3% accuracy under the homoglyph attack, RAID 2024 |
| 5 |  ZeroGPT | 7 | 2024 | Free users who want several outside tests | 77 of 104 AI samples undetected, Perkins 2024 |
| 6 |  Turnitin | 5 | 2026 | Institutions licensing Turnitin Originality | 1 of 5 records is Turnitin's own statements; 40 of 40 fully AI papers missed, Van Vlasselaer 2026 |
| 7 |  Writer | 5 | 2023 | Readers checking what four 2023 studies found for a withdrawn tool | Tool sunset 22 December 2025; 14% of 42 human papers called AI, Walters 2023; 19 of 36 AI documents undetected, Weber-Wulff 2023 |
| 8 |  Pangram | 4 | 2026 | Readers who want 2025 and 2026 results | 65% strict accuracy on 40 fully AI papers, Van Vlasselaer 2026 |
| 9 |  GLTR | 4 | 2024 | Readers who want to see the evidence word by word | 99.3% false positive rate at threshold 0.5, RAID 2024 |
| 10 | DetectGPT | 4 | 2023 | Readers who want the paraphrase results on a zero-shot method | 7.6% still detected after one DIPPER paraphrase at 1% false positives, Krishna 2023; 13 of 36 AI documents undetected, Weber-Wulff 2023 |
| 11 |  Content Detector AI | 3 | 2025 | Readers who want three outside readings of a free tool that now redirects | Lowest overall share of 16 in Walters 2023 (63% of 126, 27% uncertain); the page redirected to a Moxby Mod by 9 November 2026 |
| 12 |  Sapling | 3 | 2024 | Developers who want an API | 28% of 126 documents returned uncertain, Walters 2023 |
| 13 |  BrandWell | 3 | 2023 | Reading the record of the tool tested as Content at Scale | 36 of 36 AI documents undetected, Weber-Wulff 2023; detector page redirects to a Moxby Mod since 29 August 2026 |
| 14 |  Binoculars | 2 | 2025 | Researchers who want an open method's RAID and DAMAGE results side by side | 0.6 accuracy on open non-chat models under a repetition penalty (RAID 2024); 28.23% true positive rate after 19 humanizers (DAMAGE 2025) |
| 15 |  Hive | 2 | 2025 | Image checks with an outside record | The one record scored images, not text; 67.56% of AI images detected under high Glaze, Ha and Passananti 2024 |
| 16 |  QuillBot | 2 | 2025 | Students inside the QuillBot suite who want the two outside readings | Both records print scores, not rates: 90 AI on its own paraphrase of one text, Kar 2024; mean AI score 95.4% on DeepThink text falling to 58% after a humanized paraphrase, Alshammari and Rao 2025 |
| 17 |  RADAR | 2 | 2025 | Researchers who want an adversarially trained open method's RAID and DAMAGE results side by side | 3.48% of RAID's human documents flagged at threshold 0.5; 3.33% true positive rate at 5% FPR on student essays and a 51.87% default-threshold false positive rate in DAMAGE 2025 |
| 18 |  GPTKit | 2 | 2024 | Readers who want the zero false positive count and the lowest post-attack accuracy side by side | 0 of 114 human texts flagged, Orenstrakh 2023; 4.5% accuracy on AI text after adversarial techniques, the lowest of seven, Perkins 2024 |
| 19 |  Winston AI | 2 | 2024 | Users checking text and images | 12 of 36 AI documents undetected, Weber-Wulff 2023 |
| 20 |  Writefull | 2 | 2023 | Readers who meet a Writefull percentage in a 2023 or 2024 file | 23 of 54 correct, rank 12 of 14, Weber-Wulff 2023; 21.6% true positive rate, Pegoraro 2023; the tool is discontinued |
| 21 |  Sightengine | 1 | 2025 | Image moderation teams who want the one outside result on the API | 50 clean and 50 attacked AI images, labelled preliminary by the authors: 0.66 (accuracy) clean, 0.00 after the attack, RAID image 2025 |
| 22 |  AI or Not | 1 | 2024 | Image checks with an outside record and a free text allowance | The one record scored images under the Optic name: 24.47% of 280 human artworks flagged, Ha and Passananti 2024 |
| 23 | Fast-DetectGPT | 1 | 2024 | Researchers who want the RAID result for an open zero-shot method | 23.2% of human documents flagged at threshold 0.5; 0.6 accuracy on open non-chat models under sampling with a repetition penalty, RAID 2024 |
| 24 |  Illuminarty | 1 | 2024 | Readers who want the one outside result on an image detector | 67.40% of 280 human artworks flagged as AI, Ha and Passananti 2024; the creator said the model had not been updated for six months |
| 25 |  Undetectable AI | 1 | 2024 | Readers who want the one outside reading of a humanizer vendor's detector | A single 500-word text and three paraphrases, all read as 100 AI, Kar 2024; no rate |
| 26 |  Wordtune | 1 | 2024 | Readers who want the one outside reading of a withdrawn free tool | One ChatGPT article and three paraphrases, no human text, Kar 2024; the page has returned 404 since July 2025 |
| 27 |  AI Writing Check | 1 | 2023 | Readers who meet a 2023 result from a nonprofit tool taken down the same year | One averaged rate across seven tools, no per-tool figure in the text; the tool has been withdrawn since August 2023 |
| 28 |  Check For AI | 1 | 2023 | Readers who want the Weber-Wulff result for a tool sunset the same year | 32 of 54 correct, rank 6 of 14; 1 false accusation in 18; sunset notice since July 2023 |
| 29 |  Compilatio | 1 | 2023 | Institutions reading a 2023 result on a European LMS tool | 11 of 36 AI documents undetected and 7 of 9 QuillBot-paraphrased texts missed, Weber-Wulff 2023 |
| 30 |  GPT Radar | 1 | 2023 | Readers who meet a GPT Radar percentage from a 2023 or 2024 check | 36% of 84 AI papers called human, the highest miss rate of 16 tied with Writer; the domain has redirected since August 2026 |
| 31 |  Grammica | 1 | 2023 | Readers who want the one outside reading of a free tool removed in 2026 | 17% of 84 AI papers called human; the detector was removed in the site relaunch of January 2026 |
| 32 |  IvyPanda | 1 | 2023 | Readers who want a free colour-coded check with one outside reading | 23% of 126 documents returned as uncertain; the vendor calls the tool a beta and publishes no figure |
| 33 |  PlagiarismCheck | 1 | 2023 | Readers who want the one outside reading of an LMS-integrated plagiarism suite's detector | 21 of 54 correct, rank 13 of 14; 19 of 36 AI documents undetected; the vendor knew the documents were a test suite, Weber-Wulff 2023 |
| 34 |  Scribbr | 1 | 2023 | Students who want unlimited free checks | 69% correct on 42 GPT-4 papers, Walters 2023 |
| 35 |  SEO.ai | 1 | 2023 | Readers who want the one outside reading of a free no-login tool with a 98.4% claim | 12% of 42 human papers called AI, the third-highest false positive rate of 16; 24% of 126 documents uncertain |
Eleven records added in the sixth wave on 9 November 2026 changed eight counts and brought three cards in: Kar 2024 (six cards, one text), Alshammari and Rao 2025 (four cards, 49 answers), DAMAGE 2025 (GPTZero and Pangram), Epoch AI 2026 (Pangram, GPTZero, Originality.ai), Ibrahim 2023 (the GPT-2 Output Detector) and RAID image 2025 (Hive, Sightengine) lifted GPTZero to 16, the GPT-2 Output Detector to 12, Copyleaks to 9, Originality.ai and ZeroGPT to 7, Pangram to 4, Sapling to 3 and Hive to 2, and gave QuillBot (2), Undetectable AI (1) and Sightengine (1) their first records, so 21 of 42 cards qualified at that version; on 10 November 2026 the ninth wave added nine cards, four with records (Content Detector AI 3, GPTKit 2, Fast-DetectGPT 1, PlagiarismCheck 1), so 25 of 51 cards qualified at that version; later on 10 November 2026 the tenth wave added twelve cards, four with records (Binoculars 2, Writefull 2, Wordtune 1, Check For AI 1), so 29 of 63 cards qualified at that version; the eleventh wave, the same day, added nine cards, six with records (RADAR 2, GPT Radar 1, Grammica 1, IvyPanda 1, SEO.ai 1, AI Writing Check 1) and raised the GPT-2 Output Detector to 13 with Rashidi 2023, so 35 of 72 cards now qualify. Three further records in the eleventh wave count for no card: ASVspoof 2021, ADD 2023 and FaceForensics++ 2019 score anonymised teams and retrained research methods, and the Originality.ai compilation is a vendor statement outside every count. Three records in the same wave count for no card: ASVspoof 5, SAFE 2025 and DFDC 2020 score anonymised challenge entrants, the Vanderbilt 2023 statement is listed on the Turnitin card as a statement rather than a test, and the GPTZero and Pangram vendor benchmarks are the vendors' own measurements, listed on those two cards as benchmark lines outside the count.
One of ten rankings in the rankings hub, beside false positive rates and languages supported. The home board answers how many AI detectors have an independent study behind them?
Which AI detectors should you check first?
- Ifyou want the most outside results: GPTZero, 12 records.
- Ifyou want a free monthly allowance with outside tests: ZeroGPT, 5 records.
- Ifyou want a free, open-source tool with an outside record: the GPT-2 Output Detector, 9 records, or GLTR, 3 records.
- Ifyou want an image detector with an outside record: Hive or AI or Not, 1 record each, the only two on image detection (studies by content type).
- Ifyou are weighing one of the 37 cards with no study record: run your own test first.
What did each study find, tool by tool?
GPTZero
- Verdict
- Fits readers weighing eleven disagreeing studies.
- Value
- 16 study records; the count rose from 11 on 9 November 2026 when Pegoraro 2023 was added (true positive rate 27.3%, true negative rate 93.5%, Table I); the count rose from 12 to 16 on 9 November 2026 when Kar 2024 (97% AI on one ChatGPT 3.5 article, 29 after a QuillBot paraphrase), Alshammari and Rao 2025 (mean AI score 100% on 49 DeepSeek-V3 answers, 52% after a humanized paraphrase), DAMAGE 2025 (60.04% true positive rate on humanized essays at 5% FPR) and Epoch AI 2026 (0 of 495 human passages flagged, 32 of 297 style-imitation passages missed) were added.
- Best for
- Educators comparing vendor claims with outside tests.
- Main limitation
- Weber-Wulff 2023 recorded 9 false accusations in 18 human-written documents, the highest of 14 tools (arXiv 2306.15666, read 8 November 2026); Orenstrakh 2023 recorded 52 false positives in 114 human coursework submissions (arXiv 2307.07411, read 9 November 2026).
ZeroGPT
- Verdict
- Fits free users who want outside records.
- Value
- 7 study records (fifth by the tie rule: Originality.ai, also at 6, has a newer record); the count rose from 5 on 9 November 2026 when Pegoraro 2023 was added (true positive rate 45.7%, the highest of 12 online tools, true negative rate 92.2%, Table I); the count rose from 6 to 7 on 9 November 2026 when Kar 2024 (95.03% AI on one ChatGPT 3.5 article, 32.31 after a QuillBot paraphrase) was added.
- Best for
- Students checking drafts at no cost.
- Main limitation
- Perkins 2024 left 77 of 104 AI samples undetected (arXiv 2403.19148, read 8 November 2026).
- Studies
- Weber-Wulff 2023, Liang 2023, Walters 2023, Perkins 2024, RAID 2024.
Turnitin
- Verdict
- For institutions licensing Turnitin Originality.
- Value
- 5 study records, including 1 record of Turnitin's own statements.
- Best for
- Universities weighing false accusations above missed AI text.
- Main limitation
- Perkins 2024 found accuracy on AI text fell from 50% to 7.9% after adversarial edits, the largest drop of seven tools (arXiv 2403.19148, read 8 November 2026); Van Vlasselaer 2026 scored all 40 fully AI-generated papers under 20% (DOI 10.1007/s40979-026-00226-w, read 9 November 2026).
GPT-2 Output Detector
- Verdict
- Fits researchers who want a free, open baseline with the longest test record after GPTZero.
- Value
- 13 study records; the count rose from 12 to 13 on 10 November 2026 when Rashidi 2023 was added (up to 8.69% of 1,600 real journal abstracts per five-year range scored above 50% AI probability, no AI text scored); the count rose from 9 on 9 November 2026 when Pegoraro 2023 was added (run as the Solaiman et al. RoBERTa detector, true positive rate 7.2%, true negative rate 96.4%, Table I); the count rose from 10 to 12 on 9 November 2026 when Ibrahim 2023 (16 of 120 ESL essays below a 95% originality cut, 10 of 120 ChatGPT essays above a 5% cut) and Alshammari and Rao 2025 (mean AI score 8.15% on 49 DeepSeek-V3 answers) were added.
- Best for
- Readers who want to run the same model the papers tested, locally and at no cost.
- Main limitation
- Perkins 2024 left 81 of 104 AI samples undetected and the model card warns against using the tool as a ChatGPT detector (arXiv 2403.19148; model card, read 9 November 2026); Jabarian and Imas 2025 recorded a false positive rate of 0.9503 at threshold 0.5 on 1,992 human passages for the RoBERTa-base baseline (NBER w34223, read 9 November 2026).
Copyleaks
- Verdict
- Fits LMS schools.
- Value
- 9 study records; the count rose from 6 on 9 November 2026 when Pegoraro 2023 was added (true positive rate 22.9%, true negative rate 92.1%, Table I); the count rose from 7 to 9 on 9 November 2026 when Kar 2024 (100% AI on one ChatGPT 3.5 article, 0 after a QuillBot paraphrase) and Alshammari and Rao 2025 (mean AI score 98.8% on 49 DeepSeek-V3 answers, 71% after a humanized paraphrase) were added.
- Best for
- Canvas, Moodle, Blackboard or D2L deployments.
- Main limitation
- Perkins 2024 recorded 5 false accusations in 10 human samples, the highest of seven tools (arXiv 2403.19148, read 8 November 2026); Orenstrakh 2023 flagged all 10 Spanish human papers as AI (arXiv 2307.07411, read 9 November 2026).
Originality.ai
- Verdict
- Fits publishers checking long-form text.
- Value
- 7 study records; the count rose from 5 on 9 November 2026 when Pegoraro 2023 was added (true positive rate 7.6%, true negative rate 95.0%, Table I); the count rose from 6 to 7 on 9 November 2026 when Epoch AI 2026 (19 of 495 pre-2022 human passages flagged, 3.84%; 53 of 297 style-imitation passages missed) was added.
- Best for
- Content teams scanning articles before publication.
- Main limitation
- RAID 2024 recorded 9.3% accuracy under the homoglyph attack, down from 85.0% with no attack (arXiv 2405.07940, read 8 November 2026); Jabarian and Imas 2025 recorded a false negative rate of 0.4242 on Claude Opus 4 text at the 0.9 threshold (NBER w34223, read 9 November 2026).
Pangram
- Verdict
- Fits readers who want the newest outside results, from 2025 and 2026.
- Value
- 4 study records; the count rose from 2 to 4 on 9 November 2026 when DAMAGE 2025 (the vendor's own research models, 98.26% true positive rate on humanized essays at 5% FPR) and Epoch AI 2026 (0 of 495 human passages flagged, 0 of 297 plain-prompt passages missed, 30 of 297 style-imitation passages missed) were added.
- Best for
- Readers weighing a tool tested on long academic papers and on 1,992 pre-2020 human passages.
- Main limitation
- Van Vlasselaer 2026 recorded 65% strict accuracy on 40 fully AI-generated papers (97.5% inclusive), and the authors' run on 1,163 theses without ground truth flagged 45.5% (DOI 10.1007/s40979-026-00226-w, read 9 November 2026).
Winston AI
- Verdict
- Fits users checking text and images together.
- Value
- 2 study records.
- Best for
- Writers and small teams checking mixed content.
- Main limitation
- Weber-Wulff 2023 left 12 of 36 AI documents undetected, mostly edited or paraphrased texts (arXiv 2306.15666, read 8 November 2026).
- Studies
- Weber-Wulff 2023, RAID 2024.
Sapling
- Verdict
- Fits developers who want an API.
- Value
- 3 study records; the count rose from 2 to 3 on 9 November 2026 when Kar 2024 (100% AI on one ChatGPT 3.5 article and on all three paraphrases) was added.
- Best for
- Teams integrating detection into their own software.
- Main limitation
- Walters 2023 returned uncertain on 28% of 126 documents and scored 33% correct on 42 GPT-4 papers (DOI 10.1515/opis-2022-0158, read 8 November 2026).
- Studies
- Liang 2023, Walters 2023.
Writer
- Verdict
- Fits readers who want to know what four 2023 studies found for a tool that no longer exists.
- Value
- 5 study records; the count rose from 4 on 9 November 2026 when Pegoraro 2023 was added (true positive rate 6.9%, the lowest of 20 scored methods, true negative rate 94.5%, Table I).
- Best for
- Readers tracing a 2023 result to a detector that was withdrawn on 22 December 2025 (archived detector page, read 9 November 2026).
- Main limitation
- Walters 2023 found Writer called 14% of 42 human-written papers AI, tied for the highest false positive rate of 16 detectors (DOI 10.1515/opis-2022-0158, read 9 November 2026); Weber-Wulff 2023 left 19 of 36 AI documents undetected (arXiv 2306.15666, read 9 November 2026).
- Studies
- Weber-Wulff 2023, Walters 2023, Elkhatat 2023, Chaka 2023.
GLTR
- Verdict
- Fits readers who want to see the evidence word by word.
- Value
- 4 study records; the count rose from 3 on 9 November 2026 when Pegoraro 2023 was added (run as the Gehrmann et al. method, true positive rate 32.0%, true negative rate 98.4%, Table I).
- Best for
- Teaching what a detector reads, since every word is coloured by its rank.
- Main limitation
- RAID 2024 recorded a 99.3% false positive rate at the default threshold of 0.5, the highest of 12 detectors (arXiv 2405.07940, read 9 November 2026); Orenstrakh 2023 recorded 20 false positives in 114 through the authors' z-score wrapper (arXiv 2307.07411, read 9 November 2026).
- Studies
- RAID 2024, Chaka 2023, Orenstrakh 2023.
AI or Not
- Verdict
- Fits readers who want an image detector with an outside record and a free monthly text allowance.
- Value
- 1 study record, on image detection, under the tool's 2023 name Optic.
- Best for
- Checking images where a 2024 outside result exists; the record scored no text.
- Main limitation
- Ha and Passananti 2024 recorded a 24.47% false positive rate on 280 human artworks and a fall in AI image detection from 98.85% to 58.00% under high-intensity Glaze (arXiv 2402.03214, read 9 November 2026).
- Studies
- Ha and Passananti 2024.
Hive
- Verdict
- Fits readers who want an image detector with an outside record.
- Value
- 2 study records, on image detection; the count rose from 1 to 2 on 9 November 2026 when RAID image 2025 (0.60 accuracy on 50 clean AI images, 0.44 after a transferable adversarial perturbation) was added.
- Best for
- Checking artwork and images; the record scored no text.
- Main limitation
- Ha and Passananti 2024 found detection of AI images fell from 96.83% to 67.56% under high-intensity Glaze, while the false positive rate rose from 0% to 3.23% on 280 human artworks (arXiv 2402.03214, read 9 November 2026).
- Studies
- Ha and Passananti 2024.
Scribbr
- Verdict
- Fits students who want unlimited free checks.
- Value
- 1 study record.
- Best for
- Students checking drafts before submission.
- Main limitation
- Walters 2023 found 69% correct on 42 GPT-4 papers against 100% on 42 GPT-3.5 papers (DOI 10.1515/opis-2022-0158, read 8 November 2026).
- Studies
- Walters 2023.
DetectGPT
- Verdict
- Fits readers who want to see what three 2023 papers did to a zero-shot method.
- Value
- 4 study records: Weber-Wulff 2023 (25 of 54 documents correct through the public demo, 0 false accusations in 18 human documents), Sadasivan 2023 (AUROC 96.5% to 59.8% after one T5 paraphrase) and Krishna 2023 (70.3% to 4.6% on 3,000 Wikipedia prompts after one DIPPER paraphrase at 1% false positives), all read 9 November 2026; the count rose from 3 on 9 November 2026 when Pegoraro 2023 was added (run as the Mitchell et al. method, true positive rate 18.1%, true negative rate 80.0%, Table I).
- Best for
- Researchers comparing zero-shot curvature detection with trained classifiers.
- Main limitation
- No hosted product: the demo at detectgpt.ericmitchell.ai returned 404 on 9 November 2026, so the method runs only from the MIT-licensed code with a source model and T5.
BrandWell
- Verdict
- Fits readers tracing the tool both 2023 studies tested as Content at Scale.
- Value
- 3 study records: Weber-Wulff 2023 (18 of 54 documents correct, rank 14 of 14, 0 false accusations in 18 human documents, 36 of 36 AI documents undetected) and Walters 2023 (71% of 126 documents correct, 10% of 42 human papers called AI), both read 9 November 2026; the count rose from 2 on 9 November 2026 when Pegoraro 2023 was added (tested as Content at Scale, true positive rate 38.4%, true negative rate 79.8%, Table I).
- Best for
- Readers who want the published record behind a renamed detector.
- Main limitation
- The BrandWell detector page has redirected to a Moxby marketplace Mod since 29 August 2026 (Wayback CDX, read 9 November 2026), so the 2023 results describe a product sold under a third name.
Compilatio
- Verdict
- Fits institutions that want the one outside result on a tool sold through Moodle, Canvas and Blackboard.
- Value
- 1 study record: Weber-Wulff 2023, 40 of 54 documents correct (74%, rank 2 of 14), 679 of 864 points on the logarithmic method, 0 false accusations in 18 human documents, 11 of 36 AI documents undetected, read 9 November 2026.
- Best for
- Universities comparing Compilatio Magister+ with Turnitin on the same study.
- Main limitation
- One record, on the free 2023 tool at 200 to 2,000 characters; the v4.6.2 model live since 1 October 2026 has no outside result (support.compilatio.net, read 9 November 2026).
Illuminarty
- Verdict
- Fits readers who want an outside result on an image detector and can read a 67.40% false positive rate beside it.
- Value
- 1 study record, on image detection: Ha and Passananti 2024, 72.65% accuracy, 67.40% false positive rate and 4.69% false negative rate on 280 human artworks and 350 AI images (Table 1), read 9 November 2026.
- Best for
- Readers comparing the three commercial image detectors the 2024 study ran side by side.
- Main limitation
- The paper reports that Illuminarty's creator said in November 2023 that the model had not been updated for six months; the free plan at $0.00 a month carries no stated allowance and the API answered with an empty page on 9 November 2026.
QuillBot
- Verdict
- Fits readers who want to see how a free detector read a single text and 49 DeepSeek answers.
- Value
- 2 study records: Kar 2024 (one 500-word ChatGPT 3.5 article read as 100% AI, 90 AI after QuillBot's own paraphrase, 100 after Grammarly and ChatGPT paraphrases, Table 1) and Alshammari and Rao 2025 (mean AI score 98.4% on 49 DeepSeek-V3 answers, 95.4% on DeepThink text, 58% after a humanized paraphrase, Tables III and V), both read 9 November 2026.
- Best for
- Students who already use QuillBot and want its detector's outside record.
- Main limitation
- Neither record prints a false positive or false negative rate; Kar 2024 scored one text and Alshammari and Rao 2025 report means of displayed percentages.
Undetectable AI
- Verdict
- Fits readers who want the one outside reading on a detector sold beside a humanizer.
- Value
- 1 study record: Kar 2024, one 500-word ChatGPT 3.5 article read as 100% AI and held at 100 AI through QuillBot, Grammarly and ChatGPT paraphrases (Table 1), read 9 November 2026.
- Best for
- Readers checking the vendor's own summary of this study on its detector page against the paper.
- Main limitation
- One text, no human text and no rate; the vendor's detector page quotes the study as "100% detection / 0% false positives" with no base, as the false positive ranking's vendor table records.
Sightengine
- Verdict
- Fits readers who want an adversarial image result rather than a clean accuracy.
- Value
- 1 study record, on image detection: RAID image 2025, 0.66 (accuracy) on 50 clean AI images, 0.00 on the same 50 after a RAID transferable adversarial perturbation and 1.00 on 50 real images at a 0.5 threshold (Table D.13), read 9 November 2026.
- Best for
- API customers weighing adversarial robustness on images.
- Main limitation
- 50 images, labelled preliminary by the authors, with no API version or date printed.
Content Detector AI
- Verdict
- Fits readers who want the outside record of a free detector whose own page is gone.
- Value
- 3 study records: Walters 2023 (63% of 126 documents correct, the lowest of 16, 27% uncertain, 7% of 42 GPT-4 papers correct, Table 4), Kar 2024 (78.26% AI on one ChatGPT 3.5 article, 64.58 after a QuillBot paraphrase, Table 1) and Alshammari and Rao 2025 (mean AI score 67.27% on 49 DeepSeek-V3 answers, mean human score 55.2% on 49 human answers, the lowest of six, Tables III and IV), read 9 November 2026.
- Best for
- Readers checking what three studies said about a tool before its address redirected.
- Main limitation
- No record prints a human-text false positive count; Walters 2023 prints shares, Kar 2024 scored one text and Alshammari and Rao 2025 report means of displayed scores and name the tool as the one with the most false positives without a count; contentdetector.ai returned 301 to a Moxby Mod on 9 November 2026 and the fields read the archived page of 11 May 2026.
- Studies
- Walters 2023, Kar 2024, Alshammari and Rao 2025
GPTKit
- Verdict
- Fits readers who want a detector that rarely accuses a human writer and misses most AI text, with both figures printed.
- Value
- 2 study records: Orenstrakh 2023 (100.00% on 114 human and 75.00% on 20 ChatGPT English submissions, 0 false positives in 114, the only detector of eight at zero, 90% to 30% after a QuillBot paraphrase, Tables 6, 8 and 9) and Perkins 2024 (mean accuracy 6% on unmanipulated AI text falling to 4.5% after adversarial techniques, 0 false accusations in 10, Tables 8 and 10), read 9 November 2026.
- Best for
- Readers who weigh a false accusation more heavily than a miss.
- Main limitation
- Perkins 2024 recorded 4.5% mean accuracy on AI text after manipulation, the lowest post-manipulation figure of seven detectors, and 6% before it (DOI 10.1186/s41239-024-00487-w, Table 8, read 9 November 2026).
- Studies
- Orenstrakh 2023, Perkins 2024
Fast-DetectGPT
- Verdict
- Fits researchers who want the one large-benchmark result for an open method they can run themselves.
- Value
- 1 study record: RAID 2024, run as F-DetectGPT with the repository defaults (GPT-Neo-2.7B scoring, GPT-J reference): 23.2% of human-written documents flagged at threshold 0.5 (Table 4); accuracy at 5% FPR of 98.6 on open chat models with greedy decoding and 0.6 on open non-chat models under sampling with a repetition penalty (Table 5); 73.6 with no attack and 71.8 under paraphrase (Table 16), read 10 November 2026.
- Best for
- Readers comparing an open zero-shot detector with the commercial APIs on the same 6 million generations.
- Main limitation
- The paper's own finding is that the method can fall from near-perfect accuracy to complete failure when the generator or decoding changes (Finding 4), and the 23.2% false positive rate is at a naive threshold the authors did not tune (arXiv 2405.07940, Table 4, read 10 November 2026).
- Studies
- RAID 2024
PlagiarismCheck
- Verdict
- Fits readers who want the outside record of a detector sold inside a plagiarism checker with six LMS integrations.
- Value
- 1 study record: Weber-Wulff 2023, 21 of 54 documents correct (39% binary accuracy, rank 13 of 14, Table 5), 0 false accusations in 18 human-written documents (Table 9) and 19 of 36 AI documents undetected (52.8%, Table 10), tested 2 to 8 May 2023, read 9 November 2026.
- Best for
- Institutions on Canvas, Moodle, Blackboard, Brightspace, Schoology or Google Classroom weighing the suite's detector against its 2023 result.
- Main limitation
- The authors note that Turnitin and PlagiarismCheck knew the documents were a test suite; the tool ranked 13th of 14 and missed more than half the AI documents (arXiv 2306.15666, Tables 5 and 10, read 9 November 2026); the vendor's July 2026 benchmark post is its own measurement and feeds no field.
- Studies
- Weber-Wulff 2023
Binoculars
- Verdict
- Fits researchers who want the two benchmark results for a method they can run themselves.
- Value
- 2 study records: RAID 2024 (0.00% of human-written documents flagged at threshold 0.5, Table 4; 79.6 accuracy at 5% FPR with no attack and 80.3 under paraphrase, Tables 15 and 16) and DAMAGE 2025 (94.15% true positive rate on AI-written student essays at 5% FPR falling to 28.23% after 19 commercial humanizers, Table 3; 5.40% of human essays flagged at the default threshold, Table 7), read 10 November 2026.
- Best for
- Readers comparing an open zero-shot detector with the commercial APIs on the same generations.
- Main limitation
- RAID prints 0.6 accuracy on open non-chat models under sampling with a repetition penalty (Table 5) and DAMAGE prints 28.23% after humanizers, so the clean-text results sit beside two collapses (arXiv 2405.07940 and arXiv 2501.03437, read 10 November 2026).
- Studies
- RAID 2024, DAMAGE 2025
Writefull
- Verdict
- Fits readers who want the two outside readings of a free tool the vendor has since withdrawn.
- Value
- 2 study records: Weber-Wulff 2023 (23 of 54 documents correct, 43% binary accuracy, rank 12 of 14, Table 5; 1 false accusation in 18 human documents, Table 9; 18 of 36 AI documents undetected, Table 10, tested March 2023) and Pegoraro 2023 (true positive rate 21.6% on ChatGPT responses and true negative rate 99.3% on human responses, Table I), read 10 November 2026.
- Best for
- Readers checking what two 2023 tests said about a tool whose route now serves a discontinuation notice.
- Main limitation
- Both records date from March 2023; the vendor's own notice says the tool no longer works as intended, and the archived page of 31 December 2024 is the source for the product fields (arXiv 2306.15666, arXiv 2304.01487, web.archive.org, 31 December 2024, read 10 November 2026).
- Studies
- Weber-Wulff 2023, Pegoraro 2023
Wordtune
- Verdict
- Fits readers who want the one study record of a tool whose page is gone.
- Value
- 1 study record: Kar 2024, 100% AI on one ChatGPT 3.5 article of under 500 words, one of five tools at 100%; 0 after a QuillBot paraphrase, 100 after Grammarly and ChatGPT paraphrases, each on one text (Table 1), read 10 November 2026.
- Best for
- Readers tracing a Wordtune percentage from 2023 or 2024.
- Main limitation
- One text, no human text and no rate; the product page returned 404 from the Wayback capture of 26 July 2025 and on 10 November 2026 (DOI 10.1177/02537176241247934, web.archive.org, 30 March 2025, read 10 November 2026).
- Studies
- Kar 2024
Check For AI
- Verdict
- Fits readers who want the one outside reading of a free nonprofit tool that closed in 2023.
- Value
- 1 study record: Weber-Wulff 2023, 32 of 54 documents correct (59% binary accuracy, rank 6 of 14, tied with ZeroGPT, Table 5), 1 false accusation in 18 human-written documents (5.6%, Table 9) and 12 of 36 AI documents undetected (33.3%, Table 10), tested 7 to 28 March 2023, read 10 November 2026.
- Best for
- Readers reading a March 2023 result against a vendor page that no longer exists.
- Main limitation
- The service was sunset between the Wayback captures of 3 June and 23 July 2023 and the host did not answer on 10 November 2026; every product field is archived (arXiv 2306.15666, web.archive.org, 27 May 2023, read 10 November 2026).
- Studies
- Weber-Wulff 2023
RADAR
- Verdict
- Fits researchers who want an open detector's two benchmark results, one of them a collapse.
- Value
- 2 study records: RAID 2024 (3.48% of human-written documents flagged at threshold 0.5, Table 4; 70.9 with no attack and 67.3 under paraphrase at 5% FPR, Tables 6 and 16) and DAMAGE 2025 (3.33% true positive rate at 5% FPR on AI-written student essays, 5.05% after 19 humanizers, Table 3; 51.87% of human essays flagged at the default threshold, Table 7), read 10 November 2026. The NIST AI 700-1 naming of RADAR-Vicuna-7B as a baseline prints no result and is a note on the card, not a record.
- Best for
- Readers comparing an adversarially trained open method with the commercial APIs on the same generations.
- Main limitation
- DAMAGE 2025 prints a 3.33% true positive rate at 5% FPR because the default threshold flags 51.87% of human essays, so the detector catches almost nothing once the false positive rate is fixed (arXiv 2501.03437, Tables 3 and 7, read 10 November 2026); RAID prints 12.9 on Python code and 44.7 across its non-English and code domains (Table 7).
- Studies
- RAID 2024, DAMAGE 2025
AI Writing Check
- Verdict
- Fits readers who want the one outside reading of a free nonprofit tool its vendor withdrew.
- Value
- 1 study record: Liang 2023, one of seven detectors whose average false positive rate on 91 human-written TOEFL essays was 61.22% and on 88 US eighth-grade essays 5.19%, accessed 15 March 2023 under the name "Quil.org (aiwritingcheck.org)"; per-tool rates appear only in a chart (arXiv 2304.02819v3, read 10 November 2026).
- Best for
- Teachers tracing a spring 2023 AI Writing Check verdict on a non-native writer's essay.
- Main limitation
- The paper prints no per-tool number in its text, so no field on the card carries a rate; the vendor's notice has read "AI Writing Check is no longer available" since August 2023 (aiwritingcheck.org, read 10 November 2026).
- Studies
- Liang 2023
GPT Radar
- Verdict
- Fits readers who want the one outside reading of a paid tool its vendor let lapse.
- Value
- 1 study record: Walters 2023, 76% of 126 documents classified correctly with 24% incorrect and no uncertain responses, 36% of the 84 AI-generated papers called human and 100% of the 42 human papers correct, tested June 25 to July 12, 2023 (DOI 10.1515/opis-2022-0158, Table 4, read 8 November 2026).
- Best for
- Publishers reading a GPT Radar verdict against the one study that scored it.
- Main limitation
- The 36% miss rate is the highest in the study, tied with Writer; the home page has redirected to neuraltext.com since the Wayback capture of 14 August 2026 and the host did not resolve on 10 November 2026 (web.archive.org, 12 July 2026, read 10 November 2026).
- Studies
- Walters 2023
Grammica
- Verdict
- Fits readers who want the one outside reading of a free no-account tool that no longer exists.
- Value
- 1 study record: Walters 2023, 86% of 126 documents classified correctly with 11% incorrect and 3% uncertain, the sixth-highest overall share of 16 and the third-highest among the ten free tools; 0 of the 42 human papers called AI and 17% of the 84 AI papers called human (DOI 10.1515/opis-2022-0158, Table 4, read 8 November 2026).
- Best for
- Readers tracing a Grammica Real or Fake percentage from 2023 to 2025.
- Main limitation
- The record names no per-tool human-paper count for Grammica, so the card carries no false positive value; the detector page's last 200 capture is dated 31 July 2025 and the address returned 307 to the home page on 10 November 2026 (web.archive.org, 31 July 2025, read 10 November 2026).
- Studies
- Walters 2023
IvyPanda
- Verdict
- Fits readers who want the one outside reading of a free colour-map tool that never gives a wrong verdict, because it gives an uncertain one instead.
- Value
- 1 study record: Walters 2023, 77% of 126 documents classified correctly and no incorrect verdict, with the remaining 23% returned as uncertain, a record shared only with Copyleaks, Turnitin and Originality.ai among the 16 (DOI 10.1515/opis-2022-0158, Table 4, read 8 November 2026).
- Best for
- Students reading a colour-coded result against the one study that scored it.
- Main limitation
- The record names no per-tool human-paper count for IvyPanda, so the card carries no false positive value; the vendor's page calls the tool "a free beta test that will be improved later" and prints no figure (ivypanda.com/tools/gpt-essay-checker, read 10 November 2026).
- Studies
- Walters 2023
SEO.ai
- Verdict
- Fits readers who want a vendor's 98.4% beside the one study that scored the tool at 72%.
- Value
- 1 study record: Walters 2023, 72% of 126 documents classified correctly with 24% uncertain, 12% of the 42 human papers called AI (the third-highest false positive rate of 16) and 92% of the 84 AI papers correct (DOI 10.1515/opis-2022-0158, Table 4, read 8 November 2026).
- Best for
- Marketers reading a free SEO.ai score against its one record.
- Main limitation
- The 98.4% claim on the detector page carries no dataset, date or method, and the same page's FAQ says AI detectors "aren't highly accurate" (seo.ai/tools/ai-content-detector, read 10 November 2026).
- Studies
- Walters 2023
How we ranked
- Field sorted
- Count of indexed studies that tested the tool
- Direction
- Higher first
- Tie rule
- Most recent study date
- Qualification
- One indexed study
- Qualified
- 35 of 72 detectors
Field sorted: the study record count on each detector card, compiled on 9 November 2026. Direction: more records rank higher. Tie rule: most recent study year, then product name alphabetically; at 5 records each, Turnitin (2026 record) ranks above Originality.ai (2025) and ZeroGPT (2024); at 2 each, Pangram (2026) ranks above Winston AI (2024) and Sapling (2023); at 3 each, GLTR (2024) ranks above DetectGPT (2023); at 2 each, BrandWell (2023) ranks above Sapling (2023) by name; at 1 each, AI or Not (2024) ranks above Hive (2024) by name and both above Compilatio and Scribbr (2023), Compilatio first by name; Writer's 4 records, all dated 2023, place it seventh on its own. Qualification rule: at least one study record; cards with zero records are listed, never ranked. 10 of 22 qualified on the morning of 9 November 2026, when the GPT-2 Output Detector and GLTR were added; 12 of 22 qualified after the eight records added later that day; 14 of 28 qualified after the six cards of the third wave, two of them (Writer and AI or Not) with records; 35 of 72 qualify after the ten cards of the fourth wave, three of them (DetectGPT, BrandWell and Compilatio) with records. Turnitin's count includes a 2026 record of its own statements, because the index files every dated record a card cites. Full rules: methodology page; every record: studies index.
Why numbers differ between sources
The studies used different samples. Jabarian and Imas 2025 used 1,992 human passages written before 2020 and 7,968 AI passages (NBER w34223); Van Vlasselaer 2026 used 160 academic papers of 4,000 words or more (DOI 10.1007/s40979-026-00226-w); Orenstrakh 2023 used 164 coursework submissions (arXiv 2307.07411), all read 9 November 2026. Weber-Wulff 2023 used 54 documents including machine-translated and paraphrased texts (arXiv 2306.15666); Walters 2023 used 126 student essays, 42 human-written (DOI 10.1515/opis-2022-0158); Perkins 2024 used 10 human and 104 adversarially edited AI samples (arXiv 2403.19148); RAID 2024 reports rates at four thresholds (arXiv 2405.07940). All read 8 November 2026. Copyleaks posts 0 false accusations in one paper and 5 of 10 in another.
Weber-Wulff 2023, the record most cards in this ranking share, tested 14 tools of which 9 now have a card; the Weber-Wulff roster question, added 9 November 2026, lists all 14 with their status at that date and the records each carded tool has gained since.
Questions about independently tested AI detectors
How many AI detectors have been independently tested?
35 of the 72 detectors in the index have at least one study record as of 9 November 2026: GPTZero, the GPT-2 Output Detector, Copyleaks, Turnitin, Originality.ai, ZeroGPT, Writer, GLTR, DetectGPT, Pangram, Winston AI, BrandWell, Sapling, AI or Not, Hive, Compilatio and Scribbr. No paper on the other 37 met the index rules.
Which AI detector has been tested the most?
GPTZero, with 12 study records: Pegoraro 2023, Weber-Wulff 2023, Liang 2023, Walters 2023, Elkhatat 2023, Perkins 2024, RAID 2024, Chaka 2023, Krishna 2023, Orenstrakh 2023, Jabarian and Imas 2025 and Van Vlasselaer 2026, the last four read 9 November 2026. The results run from 9 false accusations in 18 human documents to 0.00% at threshold 0.5, so the count measures attention, not performance. Sources read 8 November 2026.
Does more testing mean a more accurate detector?
No: the 11 studies on GPTZero include the highest false accusation rate in Weber-Wulff 2023 and one of the lowest in RAID 2024. A count says how much evidence exists to read; the false positive ranking sorts on one result column. Sources read 8 November 2026.
Why does Turnitin count 5 records when one is its own statement?
Turnitin's 5 records are 4 independent studies (Weber-Wulff 2023, Walters 2023, Perkins 2024, Van Vlasselaer 2026) plus a dated record of Turnitin's published claims, including a false positive rate under 1% on over 700,000 papers. The index files every dated record a card cites and labels this one a statement. Sources read 8 November 2026.
Which detectors have no study record in the index?
Ten cards have no study record as of the evening of 9 November 2026: QuillBot, Grammarly, Phrasly, JustDone, Proofademic, Walter Writes, AIDetectPlus, Codequiry, Was It AI and Reality Defender. Twelve had none that morning; Pangram gained two records (Jabarian and Imas 2025, Van Vlasselaer 2026) and Hive one (Ha and Passananti 2024) the same day. Proofademic states a 0.2% false positive rate on vendor-chosen data, filed on the card as a vendor claim. Four of the six cards added later on 9 November 2026 also have no record: Undetectable AI, Grammarly Authorship, Turnitin Clarity and SynthID. Seven of the ten fourth-wave cards have none: Smodin, Surfer, Quetext, TruthScan, Sightengine, Deepware and Sensity, so 37 cards have none at index version 2026.11.13. Of the nine cards added on 10 November 2026 with the eleventh wave, three have none: Ghostbuster, Hiya Deepfake Detector and McAfee Deepfake Detector.
What is the most recent study in the ranking?
Van Vlasselaer 2026, published 29 June 2026 and covering GPTZero, Pangram, Copyleaks and Turnitin (DOI 10.1007/s40979-026-00226-w, read 9 November 2026), is the most recent study in the ranking, followed by two 2025 reports, Jabarian and Imas and NIST AI 700-1. Before the records added on 9 November 2026, two 2024 studies were the most recent: RAID, covering GPTZero, ZeroGPT, Originality.ai, Winston AI, the GPT-2 Output Detector and GLTR among the index cards, and Perkins 2024, covering GPTZero, ZeroGPT, Turnitin, Copyleaks and the GPT-2 Output Detector. Every result predates GPTZero 4o, announced 18 October 2026 (gptzero.me, read 8 November 2026).
Was Liang 2023 counted for every detector tested in the paper?
Yes, Liang 2023 counts as 1 record each for GPTZero, ZeroGPT, Originality.ai, Sapling and the GPT-2 Output Detector, the five index cards among the seven detectors tested. The paper reports a 61.22% average false positive rate on 91 TOEFL essays across the seven and shows per-detector rates only in a chart. Sources read 8 November 2026.
How often is this ranking updated?
The ranking was created in November 2026 and changes whenever a study record is added, with the Updated line at the top changing to match. A paper qualifies when the authors name the detectors tested, the sample sizes and the results, per the methodology page. Figures from the index, compiled 8 November 2026.
Sources and revisions
Sources
- Weber-Wulff et al. 2023arxiv.org/pdf/2306.15666
- Liang et al. 2023arxiv.org/pdf/2304.02819v3
- Walters 2023, DOI 10.1515/opis-2022-0158degruyterbrill.com/document/doi/10.1515/o...
- Elkhatat et al. 2023, DOI 10.1007/s40979-023-00140-5link.springer.com/content/pdf/10.1007/s40...
- Perkins et al. 2024arxiv.org/pdf/2403.19148
- Dugan et al. 2024, RAIDarxiv.org/pdf/2405.07940
- Chaka 2023, DOI 10.37074/jalt.2023.6.2.12doi.org/10.37074/jalt.2023.6.2.12
- Turnitin AI writing detection FAQsguides.turnitin.com/hc/en-us/articles/284...
- GPTZero 4o announcementgptzero.me/news/introducing-gptzero-4o
- GPT-2 Output Detector model cardhuggingface.co/openai-community/roberta-b...
- Krishna et al. 2023arxiv.org/abs/2303.13408
- Orenstrakh et al. 2023arxiv.org/abs/2307.07411
- Jabarian and Imas 2025, NBER w34223nber.org/papers/w34223
- Van Vlasselaer et al. 2026, DOI 10.1007/s40979-026-00226-wdoi.org/10.1007/s40979-026-00226-w
- Ha, Passananti et al. 2024arxiv.org/abs/2402.03214
- Writer AI Content Detector, archived 9 December 2025web.archive.org/web/20251209115654/https:...
Revision log
- Ranking created
- Gpt-2 output detector and gltr added, 10 of 22 qualify
- Chaka 2023 record added, gptzero 7, copyleaks 4, gltr 2
- Eight wave 2 records added, pangram and hive join, 12 of 22 qualify, gptzero 11
- Six wave 3 cards added, writer (4 records) and ai or not (1) join, 14 of 28 qualify
- Ten wave 4 cards added, detectgpt (3 records), brandwell (2) and compilatio (1) join, 17 of 38 qualify
- Four wave 5 cards and two records added
- Eleven wave 6 records added (ibrahim 2023, kar 2024, alshammari and rao 2025, damage 2025, epoch ai 2026, asvspoof 5, safe 2025, dfdc 2020, raid image 2025, the vanderbilt statement and the vendor benchmarks)
- Eighth-wave pages linked
- Nine wave 9 cards added (imagga held)
- Twelve wave 10 cards added (optic held as a duplicate of ai or not)
- Nine wave 11 cards and five records added
