Studies digest, Q1 2027: the thirteen evaluations added since the Q4 digest, with one headline figure each
Thirteen study records joined the AI detector index on 9 November 2026 after the Q4 2026 studies digest closed at 17 records, taking the studies index to 35 records at version 2026.11.13: Pegoraro 2023 and Fleckenstein 2024 at version 2026.11.6 and eleven more at version 2026.11.7, the oldest posted on 12 June 2020 and the newest published on 15 July 2026. Their headline figures run from a Kaggle challenge in which no entrant reached 70% accuracy on 10,000 unseen deepfake videos in DFDC 2020 (ai.meta.com, read 9 November 2026) to two detectors flagging 0 of 495 pre-2022 human passages in Epoch AI 2026 (epoch.ai, read 9 November 2026). Two further register rows, AIDI-S-0020 Dalalah 2023 and AIDI-S-0022 Jiang 2024, are held unwritten because no primary copy of either paper was reachable, so the fifteen rows added to the register in the period yielded thirteen records.
Checked 9 November 2026 for the quarter ending 31 March 2027; the next scheduled recheck of every card is 8 December 2026.
This is the second quarterly studies digest. 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 thirteen pages. The Q1 2027 index update covers the cards, fields and rankings added in the same period, and the 2026 false positive evidence review gained the Epoch AI 2026 and Ibrahim 2023 rows. This page sits in the updates hub; the independently tested ranking counts the records per card.
| Record | Added at version | Headline figure | Base | Cards fed |
|---|---|---|---|---|
| Pegoraro 2023 (AIDI-S-0018) | 2026.11.6 | ZeroGPT's true positive rate of 45.7% was the highest of 12 online tools; Writer's 6.9% the lowest of 20 scored methods | ChatGPT and human question-and-answer text, Table I | GPTZero, GPT-2 Output Detector, Copyleaks, Originality.ai, ZeroGPT, Writer, GLTR, DetectGPT, BrandWell |
| Fleckenstein 2024 (AIDI-S-0019) | 2026.11.6 | Teachers, not software, were the raters; the record feeds no card | Rater groups of pre-service and in-service teachers | None |
| Ibrahim 2023 (AIDI-S-0021) | 2026.11.7 | 16 of 120 human ESL essays (13.3%) scored below a 95% originality cut and 10 of 120 ChatGPT essays (8.3%) above a 5% cut on the GPT-2 Output Detector | 120 human essays and 120 ChatGPT (GPT-3.5, January 2023) essays | GPT-2 Output Detector |
| Kar 2024 (AIDI-S-0023) | 2026.11.7 | Ten free detectors read one ChatGPT 3.5 article from 0% to 100% AI; Copyleaks fell from 100% to 0 after a QuillBot paraphrase | One 500-word text and three paraphrases, Table 1 | Copyleaks, QuillBot, ZeroGPT, GPTZero, Sapling, Undetectable AI |
| Alshammari and Rao 2025 (AIDI-S-0024) | 2026.11.7 | GPTZero's mean AI score was 100% on DeepSeek-V3 answers and 52% after a humanized paraphrase | 49 answers per set, 294 samples, Tables III to V | GPTZero, Copyleaks, QuillBot, GPT-2 Output Detector |
| DAMAGE 2025 (AIDI-S-0025) | 2026.11.7 | GPTZero's true positive rate fell from 99.73% to 60.04% after 19 humanizers at a fixed 5% false positive rate; the vendor's own DAMAGE model held 98.26% | Held-out student essays, count not printed, 1,000 bootstrap iterations, Table 3 | GPTZero, Pangram |
| Epoch AI 2026 (AIDI-S-0026) | 2026.11.7 | Pangram and GPTZero flagged 0 of 495 human passages; Originality.ai flagged 19 (3.84%); all three missed 10.10% to 17.85% of AI text imitating a named author | 495 human and 594 AI passages, default verdicts, not peer reviewed | Pangram, GPTZero, Originality.ai |
| ASVspoof 5 (AIDI-S-0027) | 2026.11.7 | The best closed-condition speech deepfake system reached an EER of 8.61% and the best open system 2.59%; 27 teams beat the organisers' baseline | 138,688 bona fide and 542,086 spoofed utterances, Track 1 evaluation set | None (no indexed product participated) |
| SAFE 2025 (AIDI-S-0028) | 2026.11.7 | The top team's balanced accuracy fell from 0.87 on raw synthetic speech to 0.66 after laundering through a car cabin, a speaker and a room | 4,200 real and 2,600 generated clips from 13 TTS models, Tables 6 to 10 | None (no indexed product participated) |
| DFDC 2020 (AIDI-S-0029) | 2026.11.7 | The winning model scored 65.18% (accuracy) on 10,000 black-box videos and none of 2,114 teams reached 70% | 128,154 clips from 960 paid actors; 35,109 models submitted | None (Kaggle teams, not products) |
| RAID image 2025 (AIDI-S-0030) | 2026.11.7 | Sightengine's accuracy on 50 AI images fell from 0.66 to 0.00 and Hive's from 0.60 to 0.44 after a transferable adversarial perturbation | 50 clean and 50 attacked AI images and 50 real images, Table D.13, labelled preliminary | Hive, Sightengine |
| Vanderbilt 2023 statement (AIDI-S-0031) | 2026.11.7 | Vanderbilt disabled Turnitin's AI indicator on 16 August 2023, reasoning that a 1% false positive rate on 75,000 papers implied about 750 wrongly flagged papers a year | An institutional statement, not a test | Turnitin (as a note, counted in no field) |
| Vendor benchmarks 2026 (AIDI-S-0032) | 2026.11.7 | GPTZero states a 0.08% false positive rate for model 4.3b on its own four-domain sets; Pangram states 70 of 781,981 (0.0090%) on the 4.1 model card | Vendor-chosen texts, 4,000 human for GPTZero and 781,981 for Pangram | GPTZero, Pangram (as vendor benchmark lines, counted in no field) |
Figures as each record prints them, with the table or section named on the record page; "cards fed" lists the cards whose study list carries a line from the record, and the two lines counted in no field are labelled.
What did the thirteen records add to the evidence?
Four things. First, the index now holds a 2026 result on a large human set: Epoch AI 2026 scored 495 passages by 99 professional authors, all published before 30 November 2022, and Pangram 3.3.2 and GPTZero model 2026-05-11-base flagged none of them, while Originality.ai Turbo 3.0.2 flagged 19, sixteen of them fiction (epoch.ai, results table, read 9 November 2026). The same insight shows where the misses are: 30, 32 and 53 of 297 AI passages written to imitate a named author's style went undetected, against at most 2 of 297 from a plain prompt. Second, the humanizer question has a measured answer for one tool: in DAMAGE 2025 GPTZero's true positive rate at a fixed 5% false positive rate fell from 99.73% to 60.04% after 19 commercial humanizers, and that 60.04% replaced the 20% Orenstrakh 2023 figure in the robustness field because the ranking takes the highest printed post-attack detection (arXiv 2501.03437, Table 3, read 9 November 2026); the paper is written by Pangram Labs, and its figures for Pangram's own research models stay in prose because the paper does not say they are the shipping product.
Third, the index gained its first voice and video records, and none of them scores an indexed card. ASVspoof 5 (arXiv 2601.03944, read 9 November 2026), SAFE 2025 (arXiv 2510.03387, read 9 November 2026) and DFDC 2020 (arXiv 2006.07397, read 9 November 2026) are challenges in which anonymised or academic teams compete, so they appear in the records-by-content-type table of studies by content type with zero indexed cards and lift no card's count; Pindrop Pulse carries an ASVspoof 5 note because the vendor is an organiser affiliation whose "99% on the ASVspoof5 dataset" claim neither paper scores. Fourth, the image side gained its second record: RAID image 2025 scored Sightengine and Hive on 50 AI images before and after a transferable adversarial perturbation, 0.66 to 0.00 and 0.60 to 0.44, results the authors label preliminary, which entered the adversarial robustness ranking as image values (arXiv 2506.03988, Table D.13, read 9 November 2026).
Which cards gained their first record?
Three. QuillBot gained two, Kar 2024 and Alshammari and Rao 2025, both printing scores rather than rates: 90 AI on its own paraphrase of a single text, and a mean AI score of 95.4% on DeepThink text falling to 58% after a humanized paraphrase (QuillBot card). Undetectable AI gained Kar 2024, the journal study its detector page summarises as "100% detection / 0% false positives"; the paper scored one AI text and three paraphrases, all read as 100 AI, and no human text, so the vendor's 0% has no base in the paper (Undetectable AI card). Sightengine gained RAID image 2025, with the 0.00 adversarial accuracy above (Sightengine card). GPTZero's count rose from 12 to 16, the GPT-2 Output Detector's from 10 to 12, Copyleaks' from 7 to 9, Pangram's from 2 to 4 and Hive's from 1 to 2; the independently tested ranking now lists 21 of 72 cards.
Which records changed a rank field, and which did not?
Three fields moved, each under a rule the ranking already stated. Pangram's false negative rate moved from 1.31% of 1,992 GPT-4.1 passages to 0 of 297 plain-prompt passages (Epoch AI 2026), and the GPT-2 Output Detector's from 33.3% to 8.3% (Ibrahim 2023), both printed miss counts under the no-complement rule (false negative rate ranking); GPTZero's robustness moved to 60.04% (DAMAGE 2025). No false positive rate moved: Epoch AI's 0 of 495 is not lower than the zeros already on the Pangram and GPTZero cards, so the figure was appended to each base under the "lowest study-sourced rate" rule (false positive rate ranking). Five records moved nothing by design: Pegoraro 2023 prints true positive and true negative rates, whose complements the rankings do not take; Kar 2024 scored one text; Alshammari and Rao 2025 reports means of displayed percentages; Ibrahim 2023's 13.3% on human essays carries the author's inverted label "false negatives" and stays in prose; and the vendor benchmarks are the vendors' own measurements, recorded with their dates and never ranked.
Which candidate studies are still unread?
Two register rows hold their numbers unwritten under the rule that a paper the index cannot read in full is held, not summarised from a write-up: 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), both closed access on 9 November 2026 through the publisher, the Elsevier API without an entitlement key, OpenAlex, Semantic Scholar, CORE and a title search. Either record is written the day a primary copy is read. The Nature paper on SynthID Text remains cited on the SynthID card as a vendor statement.
No detector score should be the only evidence in a misconduct, hiring or publishing decision, and thirteen records on 35 is still a thin shelf for 72 tools.
Questions about the Q1 2027 studies digest
How many AI detector studies were added to the index since the Q4 2026 digest?
Thirteen, all on 9 November 2026: Pegoraro 2023 and Fleckenstein 2024 at version 2026.11.6, then Ibrahim 2023, Kar 2024, Alshammari and Rao 2025, DAMAGE 2025, Epoch AI 2026, ASVspoof 5, SAFE 2025, DFDC 2020, RAID image 2025, the Vanderbilt 2023 statement and the GPTZero and Pangram vendor benchmarks at version 2026.11.7, taking the studies index from 17 to 35 records.
Which new record has the largest human sample?
Epoch AI 2026, with 495 human passages of about 500 words by 99 authors, on which Pangram and GPTZero recorded 0 false positives and Originality.ai 19 (3.84%) (epoch.ai, read 9 November 2026); the challenge records are larger but score no indexed product.
Why do three of the thirteen records feed no card?
ASVspoof 5, SAFE 2025 and DFDC 2020 are challenges in which anonymised teams or Kaggle entrants compete, and no indexed detector is a named participant, so each record counts as a voice or video record with zero indexed cards in studies by content type; Fleckenstein 2024 rated teachers, not software.
Sources
- The index data file, versions 2026.11.4 to 2026.11.7 (read 9 November 2026)
- Pegoraro et al. 2023: https://arxiv.org/abs/2304.01487 (read 9 November 2026)
- Ibrahim 2023: https://doi.org/10.1186/s40468-023-00260-2 (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)
- Masrour, Emi and Spero 2025, DAMAGE: https://arxiv.org/abs/2501.03437 (read 9 November 2026)
- Epoch AI 2026, AI detectors rarely flag human writing: https://epoch.ai/data-insights/ai-detectors-false-negatives (read 9 November 2026)
- Wang et al. 2026, ASVspoof 5 evaluation: https://arxiv.org/abs/2601.03944 (read 9 November 2026)
- Trapeznikov et al. 2025, SAFE challenge: https://arxiv.org/abs/2510.03387 (read 9 November 2026)
- Dolhansky et al. 2020, DFDC dataset: https://arxiv.org/abs/2006.07397 (read 9 November 2026)
- Meta AI, Deepfake Detection Challenge results, 12 June 2020: https://ai.meta.com/blog/deepfake-detection-challenge-results-an-open-initiative-to-advance-ai/ (read 9 November 2026)
- Eddoubi et al. 2025, RAID image benchmark: https://arxiv.org/abs/2506.03988 (read 9 November 2026)
- Vanderbilt University, 16 August 2023: https://www.vanderbilt.edu/brightspace/2023/08/16/guidance-on-ai-detection-and-why-were-disabling-turnitins-ai-detector/ (read 9 November 2026)
- GPTZero AI detection benchmarking: https://gptzero.me/news/gptzero-ai-detection-benchmarking-the-industry-standard-in-accuracy-transparency-and-fairness/ (read 9 November 2026)
- Pangram 4.1 model card: https://www.pangram.com/research/model-card/pangram-4-1 (read 9 November 2026)
Revision log: 2026-11-09: digest created at index version 2026.11.7.
Related pages: Index update, Q2 2027: nine cards added, nine fields added, one card held, three log lines, Studies digest, Q2 2027: no record added, seven records gained a card, two records still held, 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.