Studies digest, Q2 2027: no record added, seven records gained a card, two records still held
No study record joined the AI detector index between the Q1 2027 studies digest, which closed at 35 records on 9 November 2026 at version 2026.11.7, and version 2026.11.13 on 10 November 2026: the studies index still holds 35 records, every record the Q1 digest listed, and the seventh, eighth and ninth content waves added rankings, views, questions and nine detector cards rather than papers. What changed on the study side is which cards the records feed: seven existing records gained a card row on 10 November 2026 when four of the nine new detectors turned out to have been tested in papers already indexed, so Weber-Wulff 2023 now feeds 10 of its 14 tools rather than 9, Walters 2023 10 of 16 rather than 9, and Orenstrakh 2023 and Perkins 2024 6 of 8 and 6 of 7 (Weber-Wulff 2023; Walters 2023; Orenstrakh 2023; Perkins 2024, all read 9 November 2026). Two register rows, AIDI-S-0020 Dalalah 2023 and AIDI-S-0022 Jiang 2024, remain held because no primary copy of either paper was reachable, the same position as at the Q1 digest.
Checked 10 November 2026 for the quarter ending 30 June 2027; the next scheduled recheck of every card is 8 December 2026.
This is the third quarterly studies digest, and the first with no new record to digest. A digest lists what the period changed on the evidence side of the index, and a quarter in which the change is cards attaching to records rather than records arriving is still a change a reader quoting the Q1 digest should know about: the "cards fed" column of four Q1 and Q4 records is different today. The Q2 2027 index update covers the cards, fields and rankings added in the same period, and the 2026 false positive evidence review gained no row, because no record was added and no new false positive rate was printed. This page sits in the updates hub; the independently tested ranking counts the records per card.
| Record | Added at version | Card rows gained on 10 November 2026 | Figure the new card records | Cards fed after |
|---|---|---|---|---|
| Weber-Wulff 2023 (AIDI-S-0001) | 2026.11 | PlagiarismCheck | 21 of 54 documents correct (39%, rank 13 of 14); 0 false accusations in 18; 19 of 36 AI documents undetected (52.8%); the authors note the vendor knew the documents were a test suite | 10 of 14 tools |
| Walters 2023 (AIDI-S-0003) | 2026.11 | Content Detector AI (tested as ContentDetector.ai) | 63% of 126 documents correct, the lowest of 16, with 27% uncertain; 7% of 42 GPT-4 papers correct, the lowest in the study; no separate human-paper count printed | 10 of 16 tools |
| Perkins 2024 (AIDI-S-0005) | 2026.11 | GPTKit | Mean accuracy 6% on unmanipulated AI text falling to 4.5% after adversarial techniques, the lowest post-manipulation figure of seven; 0 false accusations in 10 | 6 of 7 tools |
| RAID 2024 (AIDI-S-0006) | 2026.11 | Fast-DetectGPT (run as F-DetectGPT with repository defaults) | 23.2% of human documents flagged at threshold 0.5; 73.6 accuracy at 5% FPR with no attack, 71.8 under paraphrase, 34.0 under synonym swap; 98.6 on open chat models and 0.6 on open non-chat models under sampling with a repetition penalty | 7 of 12 detectors |
| Orenstrakh 2023 (AIDI-S-0012) | 2026.11.3 | GPTKit | 100.00% on 114 human and 75.00% on 20 ChatGPT submissions; 0 false positives in 114, the only detector of eight at zero; 90% to 30% after a QuillBot paraphrase; 80% and 20% on 10 Spanish human and ChatGPT texts | 6 of 8 tools |
| Kar 2024 (AIDI-S-0023) | 2026.11.7 | Content Detector AI | 78.26% AI on one ChatGPT 3.5 article; 64.58 after a QuillBot paraphrase, 75 after Grammarly, 60.87 after a ChatGPT paraphrase | 7 of 10 tools |
| Alshammari and Rao 2025 (AIDI-S-0024) | 2026.11.7 | Content Detector AI | Mean AI score 67.27% on 49 DeepSeek-V3 answers and 59.25% after a QuillBot paraphrase; mean human score 55.2% on 49 human answers, the lowest of six; named by the authors as the tool with the most false positives, without a count | 5 of 6 tools |
Figures as each record prints them, with the table named on the record page; "cards fed after" counts the record's tested tools that now have a card. Four fields on the new cards took a value from these records under the rankings' standing rules: GPTKit's and PlagiarismCheck's false positive rates (0 of 114 and 0 of 18), Fast-DetectGPT's (23.2%), PlagiarismCheck's false negative rate (52.8%), GPTKit's and Fast-DetectGPT's robustness (30% and 71.8) and GPTKit's and PlagiarismCheck's claim gaps (plus 5.5 and plus 60.23 points). No value on any existing card moved.
What did the period change on the evidence side?
Three things, none of them a new paper. First, four of the nine cards added on 10 November 2026 attached to seven records that had carried "no card" rows for them, so those records now point at a page for every tool they name except the ones the index has not yet researched: Weber-Wulff 2023's remaining card-less tools are Check For AI, Crossplag, the OpenAI classifier (which has its own record) and Writeful GPT Detector; Walters 2023's are Crossplag, Grammica, GPT Radar, IvyPanda, the OpenAI classifier and SEO.ai; Orenstrakh 2023's are the OpenAI classifier and CheckForAI; Perkins 2024's is Crossplag; RAID 2024's are RoBERTa-Base (ChatGPT), RADAR, Binoculars and LLMDet (Weber-Wulff 2023; Walters 2023; RAID 2024). The Weber-Wulff roster question and the shut-down question carry the new counts.
Second, the new attachments produced the widest claim gap in the index and one of its clearest divided verdicts. PlagiarismCheck's page states "99,23%" of sentences correctly classified as AI across seven models in a vendor benchmark of July 2026, and Weber-Wulff 2023 scored the same tool at 21 of 54 documents in May 2023, a gap of 60.23 points that the claim vs evidence gap ranking now lists first from the top, with the note that the vendor's own page calls its figures marketing (plagiarismcheck.org/ai-detector, read 9 November 2026; arXiv 2306.15666, Table 5). GPTKit's two records pull in opposite directions: 0 of 114 human submissions flagged in Orenstrakh 2023, the only zero of eight, and 4.5% accuracy on manipulated AI text in Perkins 2024, the lowest of seven, so the card sits in the zero group of the false positive rate ranking and near the foot of adversarial robustness at once (arXiv 2307.07411, Tables 8 and 9; DOI 10.1186/s41239-024-00487-w, Table 8).
Third, the index gained its fifth open-source card with a benchmark result: Fast-DetectGPT's RAID 2024 row prints a false positive rate at four thresholds and accuracy under eleven attacks, which entered the false positive and robustness rankings under the GLTR treatment of a repository-default run (arXiv 2405.07940, Tables 4 and 16, read 10 November 2026). The open-source view now lists five cards, and the RAID record feeds seven of its twelve detectors.
Which records were revised in the period?
Record pages carry dated revisions since the Q1 digest, none of them changing a reported figure. In the seventh wave, twelve records gained sentences linking the five new rankings and the eleven questions, and the Jabarian and Imas 2025 record gained the GPT-4.1 false negative row at threshold 0.5 (Pangram 0.0131, GPTZero 0.0025, Originality.ai 0.1606 to 0.3733) that its cards' fields and the GPT-4 question cite. In the eighth wave, the Weber-Wulff 2023 record gained the four Table 5 rows the preprint prints for Check For AI (32), Crossplag (37), PlagiarismCheck (21) and Writeful GPT Detector (23) and the Table 9 and 10 false accusation and undetected counts for all 14 tools, so that the record and the roster question read from one source; fourteen records gained links to the eighth-wave views and questions. In the ninth wave, the seven records above gained their card rows and a sentence each. The Vanderbilt 2023 statement and the vendor benchmarks record are unchanged and still count toward no card.
Which candidate studies are still unread?
The same two as at the Q1 digest. Dalalah 2023 (AIDI-S-0020, DOI 10.1016/j.ijme.2023.100822) and Jiang, Hao, Fauss and Li 2024 on ETS-built detectors and GRE essays (AIDI-S-0022, DOI 10.1016/j.compedu.2024.105070) are closed access, and no primary copy was reachable on 9 November 2026 through the publisher, the Elsevier API without an entitlement key, OpenAlex, Semantic Scholar, CORE or a title search; the rule that a paper the index cannot read in full is held, not summarised from a write-up, keeps both numbers reserved. A third hold of the period is a card rather than a record: Imagga (AIDI-D-0052), planned as an image detector card, was found on 10 November 2026 to publish no AI-generated image detector and is held unpublished, as the Q2 2027 index update records. The content plan's eleventh month lists the next modality studies to be read (ASVspoof, DFDC and FaceForensics++ follow-ups), and none had been read by 10 November 2026.
No detector score should be the only evidence in a misconduct, hiring or publishing decision, and 35 records remains a thin shelf for 72 tools, five of the nine newest of which no study has tested.
The next digest is the Q3 2027 studies digest, which lists the five records added on 10 November 2026 with their figures and bases, Rashidi 2023 among them, and the same two held papers.
Questions about the Q2 2027 studies digest
How many AI detector studies were added to the index since the Q1 2027 digest?
None. The studies index holds 35 records at version 2026.11.13, the same count as at the Q1 digest; the period added nine detector cards, four of which attached to seven existing records (Weber-Wulff 2023, Walters 2023, Perkins 2024, RAID 2024, Orenstrakh 2023, Kar 2024, Alshammari and Rao 2025).
Which new cards have a study record?
Content Detector AI (3: Walters 2023, Kar 2024, Alshammari and Rao 2025), GPTKit (2: Orenstrakh 2023, Perkins 2024), Fast-DetectGPT (1: RAID 2024) and PlagiarismCheck (1: Weber-Wulff 2023); Decopy, Isgen, AI Detector Pro, Paperpal and Detecting-AI.com have none.
Did any existing figure on a card change?
No. The seven records gained card rows for tools they had already tested, and the new cards took their field values from figures the records already printed; no value on any of the 42 earlier cards moved, and no quoted study figure was corrected.
Sources
- The index data file, versions 2026.11.7 to 2026.11.10 (read 10 November 2026)
- Weber-Wulff et al. 2023: https://arxiv.org/pdf/2306.15666 (read 9 November 2026)
- Walters 2023: https://www.degruyterbrill.com/document/doi/10.1515/opis-2022-0158/pdf (read 8 November 2026)
- Perkins et al. 2024: https://doi.org/10.1186/s41239-024-00487-w (read 9 November 2026)
- Dugan et al. 2024, RAID: https://arxiv.org/pdf/2405.07940 (read 10 November 2026)
- Orenstrakh et al. 2023: https://arxiv.org/pdf/2307.07411 (read 9 November 2026)
- Kar, Bansal, Modi and Singh 2024: https://doi.org/10.1177/02537176241247934 (read 9 November 2026)
- Alshammari and Rao 2025: https://arxiv.org/abs/2507.17944 (read 9 November 2026)
- Jabarian and Imas 2025: https://www.nber.org/papers/w34223 (read 9 November 2026)
- PlagiarismCheck AI detector page and July 2026 benchmark: https://plagiarismcheck.org/ai-detector/ (read 9 November 2026)
- Dalalah 2023 (held, closed access): https://doi.org/10.1016/j.ijme.2023.100822 (attempted 9 November 2026)
- Jiang et al. 2024 (held, closed access): https://doi.org/10.1016/j.compedu.2024.105070 (attempted 9 November 2026)
Revision log: 2026-11-10: digest created at index version 2026.11.10; 2026-11-10: twelfth-wave pages linked
Related pages: AI detector census 2027: how many detectors the index holds, how many are gone, and how many have ever been tested, Index update, Q3 2027: twenty-one cards added, two fields added, two rows held, three log lines, AI detector accuracy claims vs published results, ranked by the gap, How accurate are AI detectors? Every study, indexed and Index updates: dated reviews of the AI detector data.