Studies digest, Q3 2027: five records added, one headline figure and base each, two closed-access papers still held
Five study records joined the AI detector index between the Q2 2027 studies digest, which closed at 30 records on 10 November 2026 at version 2026.11.10, and version 2026.11.13 later the same day: Rashidi 2023 on 14,400 real journal abstracts, the voice challenges ASVspoof 2021 and ADD 2023, the face-video benchmark FaceForensics++ 2019, and Originality.ai's own study compilation, carried as a vendor statement, taking the studies index to 35 records, all added in the eleventh wave at version 2026.11.12 (Rashidi 2023 record; ASVspoof 2021 record; ADD 2023 record; FaceForensics++ 2019 record; Originality.ai compilation record, each read in the primary source on 10 November 2026). One of the five scores an indexed card: Rashidi 2023 becomes the GPT-2 Output Detector's thirteenth record. 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 Q2 digest.
Checked 10 November 2026 for the quarter ending 30 September 2027; the next scheduled recheck of every card is 8 December 2026.
This is the fourth quarterly studies digest and the last of the twelve-month content plan; the 2027 census counts the whole study index by year and kind. A digest lists what the period changed on the evidence side: the five records, the figure each one prints with its base, the cards each one feeds, and what the tenth and twelfth waves did to records already indexed. The Q3 2027 index update covers the cards, fields and rankings of the same period, and the 2026 false positive evidence review gained no row, because the one new human-text rate, Rashidi's, sits on a card whose lowest printed rate is already 0 of 10. This page sits in the updates hub; the independently tested ranking counts the records per card.
| Record | Register ID | Kind | Headline figure, as the record prints it | Base | Cards fed |
|---|---|---|---|---|---|
| Rashidi 2023 | AIDI-S-0033 | Independent | Up to 8.69% of a five-year range's real abstracts scored above 50% AI probability (1995 to 1999) and up to 5.13% above 90% (2000 to 2004), falling to 0.75% above 90% for 2020 to 2023; 10.24% of NEJM abstracts above 90% | 14,400 abstracts from five journals, 1980 to February 2023, 1,600 per range; no AI text scored | 1 (GPT-2 Output Detector) |
| ASVspoof 2021 | AIDI-S-0034 | Challenge | Best deepfake-task detector (team T23) at 15.64% equal error rate against 22.38% for the best organiser baseline; best logical access detector at 1.32% EER and min t-DCF 0.2177 | 533,928 compressed DF trials; 148,176 LA trials; 54 teams | 0 |
| ADD 2023 | AIDI-S-0035 | Challenge | Best of 49 detection submissions (B01) at a weighted equal error rate of 12.45% over two rounds against an average of 49.94%; best manipulation-region system 67.13% against a 42.25% baseline | Track 1.2 test sets of 80,000 real and 31,976 fake utterances (round 1) and 87,500 and 30,977 (round 2), Mandarin | 0 |
| FaceForensics++ 2019 | AIDI-S-0036 | Challenge | Best forgery detector (XceptionNet on a tracked face crop) at 99.26% (accuracy) on raw frames, 95.73% lightly compressed and 81.00% heavily compressed; 204 human observers averaged 68.69%, 66.57% and 58.73% | 140 test videos from 1,000 pristine videos and four manipulations each | 0 |
| Originality.ai compilation | AIDI-S-0037 | Vendor or lab | 5.0 and 5.0 Max stated at "99%+" (accuracy) with no test-set size; 17 third-party studies restated at one headline figure each, 83% to 100% | The vendor's accuracy page and meta-analysis post dated 7 November 2026; the index checked two of the 17 against its own records | 0 (a vendor benchmarks line on the Originality.ai card, never a study row) |
Figures as each record prints them, with the table or section named on the record page; "cards fed" counts the indexed cards the record scores. The GPT-2 Output Detector's false positive field keeps its lowest printed value (0 of 10, Perkins 2024) under the ranking's rule, and its base gained the Rashidi figure; no other field moved on any card from these five.
What did the period change on the evidence side?
Three things. First, the index gained its first study-sourced human-text rate at scale on scientific writing: Rashidi 2023 is 14,400 abstracts against the 100 of Gao 2023 and the 160 papers of Van Vlasselaer 2026, and it prints the rate at five thresholds by five-year range and by journal, so a reader can see the detector flag more of the older abstracts than the newer ones (DOI 10.1016/j.jpi.2023.100342, Discussion, Fig. 2 and Table 1, read 10 November 2026; Gao 2023 record; Van Vlasselaer 2026 record). The record scored no AI text, so it carries no false negative figure and no claim gap input.
Second, the voice and video shelf doubled in records without gaining a scored card. ASVspoof 2021, ADD 2023 and FaceForensics++ 2019 join ASVspoof 5, SAFE 2025 and DFDC 2020 as challenge and benchmark records whose systems are research teams or retrained methods, not products; no card in the index is a named participant in any of the six, and the studies by content type ranking still lists the voice, video and music cards at zero records (arXiv 2210.02437; arXiv 2305.13774; arXiv 1901.08971, all read 10 November 2026). The Pindrop vs Resemble Detect comparison, added in the twelfth wave, sets two voice vendors beside that absence.
Third, the index gained its fourth vendor-statement record and kept it apart from the studies. The Originality.ai compilation restates one headline figure per study for 17 studies, including 97% on Walters 2023 where the paper prints 98%, and the record sits on the card's vendor benchmarks line with source "vendor" so that the claim vs evidence gap ranking can set the vendor's claim against the five index records that scored the product; its claim-gap inputs are noted on that ranking and not computed (meta-analysis post, read 10 November 2026; Walters 2023 record).
Which records gained cards in the period?
Seven existing records gained a card row across the tenth and eleventh waves, and an eighth a note, because ten of the 21 new cards had been tested in papers already indexed or named in them: Weber-Wulff 2023 gained Check For AI and Writefull and now feeds 12 of its 14 tools; Walters 2023 gained GPT Radar, Grammica, IvyPanda and SEO.ai and now feeds 14 of 16; Liang 2023 gained AI Writing Check (6 of 7); Pegoraro 2023 gained Writefull; Kar 2024 gained Wordtune; RAID 2024 and DAMAGE 2025 gained Binoculars and RADAR; and NIST AI 700-1 names RADAR as a baseline with no printed result, carried as a note rather than a record (Weber-Wulff 2023 record; Walters 2023 record; RAID 2024 record; NIST AI 700-1 record). The Weber-Wulff roster question now links 12 of 14 tools, with Crossplag and the OpenAI classifier the two without a card. In the twelfth wave no record gained a row; the four questions and four views added that day cite the records they rest on, and the music detectors question records that no study in the index scores AI-generated music.
Which candidate studies are still unread?
The same two as at the Q2 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, no primary copy was reachable on 9 November 2026, and the rule that a paper the index cannot read in full is held, not summarised from a write-up, keeps both numbers reserved and both rows with Kevin. The twelve-month plan's 29 study candidates are otherwise read: 35 records stand, 27 on text, 4 on voice, 2 on image and 2 on video by the hub's modality labels on 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, 37 of which no study has tested.
Questions about the Q3 2027 studies digest
How many AI detector studies were added to the index since the Q2 2027 digest?
Five, all on 10 November 2026 at version 2026.11.12: Rashidi 2023, ASVspoof 2021, ADD 2023, FaceForensics++ 2019 and the Originality.ai vendor compilation, taking the studies index from 30 to 35 records.
Which new record scores a card in the index?
Rashidi 2023 alone, as the GPT-2 Output Detector's thirteenth record (up to 8.69% of 1,600 real abstracts in a range above a 50% score); the three challenges name no indexed product and the vendor compilation sits on the Originality.ai card's vendor benchmarks line.
Which candidate papers are still held?
Dalalah 2023 (AIDI-S-0020) and Jiang 2024 (AIDI-S-0022), both closed access with no primary copy reachable on 9 November 2026; the index does not summarise a paper from a write-up, so both register IDs stay reserved.
Sources
- The index data file, versions 2026.11.10 to 2026.11.13 (read 10 November 2026)
- Rashidi et al. 2023: https://doi.org/10.1016/j.jpi.2023.100342 (read 10 November 2026)
- Liu et al. 2023, ASVspoof 2021: https://arxiv.org/abs/2210.02437 (read 10 November 2026)
- Yi et al. 2023, ADD 2023: https://arxiv.org/abs/2305.13774 (read 10 November 2026)
- Rössler et al. 2019, FaceForensics++: https://arxiv.org/abs/1901.08971 (read 10 November 2026)
- Originality.ai meta-analysis post: https://originality.ai/blog/ai-detection-studies-round-up (read 10 November 2026)
- Walters 2023: https://www.degruyterbrill.com/document/doi/10.1515/opis-2022-0158/pdf (read 8 November 2026)
- Weber-Wulff et al. 2023: https://arxiv.org/pdf/2306.15666 (read 9 November 2026)
Revision log: 2026-11-10: digest created at index version 2026.11.13 (twelfth wave).
Related pages: 2026 review: every published AI detector false positive figure in the index, Index update, Q4 2026: twelve cards added, five fields added, four corrections filed, AI image, voice, video and code detectors ranked by independent studies, How accurate are AI detectors? Every study, indexed and Index updates: dated reviews of the AI detector data.