Studies digest, Q4 2026: the nine evaluations added since launch, with one headline figure each
Nine study records joined the AI detector index between the launch build of 8 November 2026 and 9 November 2026, taking the studies index from 8 records to 35, by the index data file at version 2026.11.13: Chaka 2023 at version 2026.11.2 and eight more at version 2026.11.3, the oldest first posted on 17 March 2023 and the newest published on 29 June 2026. Their headline figures run from a watermark detector's true positive rate falling from 99.3% to 9.7% under recursive paraphrase in Sadasivan 2023 (arXiv 2303.11156, read 9 November 2026) to a false positive rate of 0.0000 on 1,992 human passages for Pangram in Jabarian and Imas 2025 (NBER w34223, read 9 November 2026). Every record was read in the primary paper, each figure below carries its base, and every figure was checked against its record on 9 November 2026.
This is the first quarterly studies digest and it covers the launch quarter. A digest lists the records added in the period with one headline figure each, the base behind the figure, and the cards the record feeds, so a reader can see what the new evidence changed without opening nine pages. The Q4 2026 index update covers the cards and fields added in the same period, and the 2026 false positive evidence review sets every false positive figure from all 35 records in one table. This page sits in the updates hub; the independently tested ranking counts the records per card, and the AI detector register shows every card they feed.
| Record | Added at version | Headline figure | Base | Cards fed |
|---|---|---|---|---|
| Chaka 2023 (AIDI-S-0009) | 2026.11.2 | Copyleaks classified 15 of 21 chatbot responses as AI-generated, the most of five tools; GPTZero classified 30 of 30 machine-translated responses as human | 21 English responses from ChatGPT, YouChat and Chatsonic; 30 translations of 6 ChatGPT responses | GPTZero, Copyleaks, GLTR, Writer |
| Sadasivan 2023 (AIDI-S-0010) | 2026.11.3 | Recursive paraphrasing cut a watermark detector's true positive rate from 99.3% to 9.7% at a 1% false positive rate; the RoBERTa-Large GPT-2 detector fell from 1.0 to 0.925 after one paraphrase | 1,000 AI and 1,000 human passages of about 300 tokens | GPT-2 Output Detector |
| Krishna 2023 (AIDI-S-0011) | 2026.11.3 | One DIPPER paraphrase cut DetectGPT from 70.3% to 4.6% and GPTZero from 13.9% to 1.2% detection at a 1% false positive rate | 3,000 prompts of 300 tokens on GPT-2 XL text | GPTZero |
| Orenstrakh 2023 (AIDI-S-0012) | 2026.11.3 | GPTZero flagged 52 of 114 human-written English submissions as AI; Copyleaks flagged 1 and all 10 Spanish human papers | 164 computing coursework submissions, 124 human and 40 ChatGPT | GPTZero, Copyleaks, Originality.ai, GPT-2 Output Detector, GLTR |
| Gao 2023 (AIDI-S-0013) | 2026.11.3 | The GPT-2 Output Detector separated real from ChatGPT abstracts with an AUROC of 0.94; blinded reviewers caught 68% of generated abstracts and wrongly flagged 14% of real ones | 50 ChatGPT and 50 real medical abstracts | GPT-2 Output Detector |
| Jabarian and Imas 2025 (AIDI-S-0014) | 2026.11.3 | False positive rates of 0.0000 for Pangram, 0.0011 for Originality.ai and 0.0071 for GPTZero at threshold 0.5, against 0.9503 for RoBERTa | 1,992 human passages written before 2020 and 7,968 AI passages from four 2025 models | Pangram, Originality.ai, GPTZero, GPT-2 Output Detector |
| Van Vlasselaer 2026 (AIDI-S-0015) | 2026.11.3 | Turnitin classified 100% of 40 fully AI-generated papers as human, GPTZero 70%, Copyleaks 75%; Pangram reached 97.5% inclusive accuracy; all four classified all 40 human papers correctly | 160 academic papers in four categories of 40, plus 1,163 theses without ground truth | GPTZero, Pangram, Copyleaks, Turnitin |
| NIST AI 700-1 2025 (AIDI-S-0016) | 2026.11.3 | 348 detector submissions from 11 teams produced AUC values from about 0.4 to 1.000, the weakest pairs at 0.3396, below chance | Summaries in three test sets of 104, 530 and 1,683 | GPT-2 Output Detector (named as a baseline; no result under its name) |
| Ha and Passananti 2024 (AIDI-S-0017) | 2026.11.3 | Hive classified images with 98.03% accuracy, 0% false positives and 3.17% false negatives; Optic (now AI or Not) flagged 24.47% of human artworks; Illuminarty 67.40% | 280 human artworks and 350 AI-generated images from five generators | Hive, AI or Not |
Figures as each paper prints them, read 8 and 9 November 2026 in the primary source linked on each record; a card is listed as fed when the record names the tool and prints a result for it, except NIST, which names the GPT-2 detector as a baseline and anonymises every team.
What did the nine records add to the evidence?
Three things the launch set lacked. First, large bases: the eight launch records used at most 126 documents outside RAID, and the quarter added 1,992 human passages in Jabarian and Imas 2025, 164 submissions in Orenstrakh 2023, 160 papers in Van Vlasselaer 2026 and 630 images in Ha and Passananti 2024, so three records now measure false positives on more than 100 human texts, RAID 2024 (14,971 documents), Jabarian and Imas 2025 (1,992 passages) and Orenstrakh 2023 (114 English submissions), where the launch set had RAID alone (false positive evidence review). Second, dates: Jabarian and Imas 2025 and Van Vlasselaer 2026 are the first records on 2025 and 2026 models, and before them the newest result in the index was RAID 2024 (independently tested ranking). Third, kinds: Ha and Passananti 2024 is the first record to score a content type other than text, Sadasivan 2023 the first to attack a watermark, Krishna 2023 the first to attack with a dedicated paraphraser, and NIST AI 700-1 the first government evaluation, with its teams anonymised.
The records also moved the rankings. Study counts rose on seven cards and brought Pangram and Hive in: 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, Pangram from 0 to 2 and Hive from 0 to 1 (corrections log). Pangram's published false positive rate moved from the vendor's 1 in 10,000 to the study figure of 0.0000 of 1,992 and the card moved from the vendor table into the false positive rate ranking. Four of the five rankings added at version 2026.11.3 are computed from these records: false negative rate from the miss counts in Walters 2023, Weber-Wulff 2023, Jabarian and Imas 2025, Ha and Passananti 2024, Chaka 2023 and Elkhatat 2023; adversarial robustness from Orenstrakh 2023, RAID 2024, Sadasivan 2023, Van Vlasselaer 2026 and Perkins 2024; studies by content type from Ha and Passananti 2024 alone; and claim vs evidence gap from the best accuracy each record prints.
Which cards gained their first record?
Three. Pangram had no record at launch and gained two on 9 November 2026, Jabarian and Imas 2025 and Van Vlasselaer 2026, which between them place it first in the false positive rate ranking's tie group on a 1,992-passage sample and show 65% strict and 97.5% inclusive accuracy on 40 fully AI papers (Pangram card). Hive gained Ha and Passananti 2024, the one image record, with 98.03% accuracy on 630 images and a fall to 67.56% detection under high-intensity Glaze (Hive card). AI or Not, carded in the third wave on 9 November 2026, arrived with the same record under its former name Optic: 90.67% accuracy, 24.47% of human artworks flagged, 1.15% of AI images missed (AI or Not card). Writer, also carded in the third wave, arrived with four records from the launch set plus Chaka 2023, all from 2023 and all predating its withdrawal on 22 December 2025 (Writer card).
Which candidate studies are still unread?
The editors' register lists further evaluations to be read in the primary source before any is indexed, under the rule that a paper the index cannot reach in full is cut rather than summarised from a write-up. Two are named on cards already: the study of ten free detectors in the Indian Journal of Psychological Medicine that Undetectable AI cites for its 0% false positive figure, and the Nature paper on SynthID Text by Dathathri et al. 2024, which is vendor-authored and is cited on the SynthID card as a vendor statement pending a decision on whether a vendor paper qualifies as a record. The next digest will list whichever of these the index reads, with the same fields.
The next digest, the Q1 2027 studies digest, written 9 November 2026, lists the thirteen records added after this one closed, including the study of ten free detectors named above, indexed as Kar 2024.
No detector score should be the only evidence in a misconduct, hiring or publishing decision, and nine records on 17 is still a thin shelf for 72 tools.
Questions about the Q4 2026 studies digest
How many AI detector studies were added in Q4 2026?
Nine, between 8 and 9 November 2026: Chaka 2023 at version 2026.11.2 and Sadasivan 2023, Krishna 2023, Orenstrakh 2023, Gao 2023, Jabarian and Imas 2025, Van Vlasselaer 2026, NIST AI 700-1 and Ha and Passananti 2024 at version 2026.11.3, taking the studies index from 8 to 35 records.
Which new study has the largest sample?
Jabarian and Imas 2025, with 1,992 human passages and 7,968 AI passages (NBER w34223, read 9 November 2026); NIST AI 700-1 scored 348 detector submissions on test sets of up to 1,683 summaries but names no product (DOI 10.6028/NIST.AI.700-1, read 9 November 2026).
Which new study is the most recent?
Van Vlasselaer 2026, published on 29 June 2026 in the International Journal for Educational Integrity and covering GPTZero, Pangram, Copyleaks and Turnitin on 160 papers (DOI 10.1007/s40979-026-00226-w, read 9 November 2026).
Sources
All read 9 November 2026.
- Chaka 2023: https://doi.org/10.37074/jalt.2023.6.2.12
- Sadasivan et al. 2023: https://arxiv.org/abs/2303.11156
- Krishna et al. 2023: https://arxiv.org/abs/2303.13408
- Orenstrakh et al. 2023: https://arxiv.org/abs/2307.07411
- Gao et al. 2023: https://doi.org/10.1038/s41746-023-00819-6
- Jabarian and Imas 2025, NBER w34223: https://www.nber.org/papers/w34223
- Van Vlasselaer et al. 2026: https://doi.org/10.1007/s40979-026-00226-w
- NIST AI 700-1: https://doi.org/10.6028/NIST.AI.700-1
- Ha, Passananti et al. 2024: https://arxiv.org/abs/2402.03214
Revision log: 2026-11-09: digest created at index version 2026.11.4; 2026-11-09: link to the Q1 2027 studies digest and the Kar 2024 record added.
Related pages: Index update, Q1 2027: fourteen cards added, four fields added, four rankings added, three log lines, Studies digest, Q1 2027: the thirteen evaluations added since the Q4 digest, with one headline figure each, 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., How accurate are AI detectors? Every study, indexed and Index updates: dated reviews of the AI detector data.