Corrections log
This corrections log records every change to a published number, field or ranking on AI Detector Index: the date, the page, what changed and the source that prompted it. The index version is 2026.11.13; the index was created on 8 November 2026 as version 2026.11 with 16 cards and 8 study records, and the log opens with that entry.
What is logged
A correction is any change to a figure a reader could have quoted: a price, a limit, a language count, a study result, a date or a rank. The entry names the page, the field, the old value, the new value and the source. A change to wording that leaves every figure intact moves the page's "Updated" date but is not logged here. The addition of a card or a study raises the index version and is logged as one line.
Log
| Date | Page | What changed | Source |
|---|---|---|---|
| 2026-11-08 | all | Index version 2026.11 created, 16 cards and 8 study records | the index data file |
| 2026-11-09 | /detectors/ | Six cards added, AIDI-D-0017 to AIDI-D-0022: GLTR, AIDetectPlus, GPT-2 Output Detector, Codequiry, Was It AI, Reality Defender; the independent testing and false positive rankings now list 10 of 22 cards | the index data file |
| 2026-11-09 | all | Index version 2026.11.2, roadmap wave 1: 6 cards, 1 study record (Chaka 2023, AIDI-S-0009, feeding GPTZero to 7 records, Copyleaks to 4 and GLTR to 2), 1 ranking (input limits), 1 view (admissions offices), 4 questions, 3 comparisons and 2 hubs (questions, compare); Copyleaks moves to third in the false positive ranking on the tie rule | the index data file and DOI 10.37074/jalt.2023.6.2.12 |
| 2026-11-09 | all | Index version 2026.11.3, roadmap wave 2: 8 study records (AIDI-S-0010 to 0017: Sadasivan 2023, Krishna 2023, Orenstrakh 2023, Gao 2023, Jabarian and Imas 2025, Van Vlasselaer 2026, NIST AI 700-1, Ha and Passananti 2024), study counts GPTZero 7 to 11, GPT-2 Output Detector 4 to 9, Copyleaks 4 to 6, Turnitin 4 to 5, Originality.ai 3 to 5, GLTR 2 to 3, Pangram 0 to 2, Hive 0 to 1; Pangram's published false positive rate moved from the vendor figure (0.01%) to Jabarian and Imas 2025 (0.0000 of 1,992); 5 rankings (data retention, studies by content type, adversarial robustness, false negative rate, claim vs evidence gap), 2 questions (how many AI detectors exist, which share an owner), 1 update (2026 false positive evidence review) and the updates hub; the independent testing and false positive rankings re-sorted on the new counts, 12 and 11 of 22 qualifying | the index data file; the eight papers, each read in the primary source on 9 November 2026 |
| 2026-11-09 | all | Index version 2026.11.4, roadmap wave 3: 6 cards (AIDI-D-0023 to 0028: Writer, Undetectable AI, AI or Not, Grammarly Authorship, Turnitin Clarity, SynthID); 2 views (open source, published false positive rate), 5 questions (shut down, stored text, machine-translated text, under 1% claims, is a detector AI), 1 comparison (GPTZero vs Pangram vs Originality.ai) and 2 updates (Q4 2026 index update, Q4 2026 studies digest); rankings re-sorted: independently tested 12 to 14 of 28 (Writer 4 records, AI or Not 1), false positive rate 11 to 12 (Writer 5.6%; Undetectable AI's vendor 0% to the vendor table), free tiers 3 to 4 (AI or Not 1,000,000 words a month), languages 12 to 13 (Undetectable AI 50+ first), data retention 5 to 9 (AI or Not and Undetectable AI 0 days, SynthID 1, Grammarly Authorship 3,650), studies by content type 14 to 16 (AI or Not first on the tie rule), false negative rate 10 to 12 (AI or Not 1.15%, Writer 40%), claim vs evidence gap 7 to 8 (AI or Not plus 8.2 points, the one image pairing); AI or Not's image false positive rate kept out of the human-text field under the Hive rule and Writer's Weber-Wulff attacked counts kept out of the robustness field under that ranking's rule; the study tables of Weber-Wulff 2023, Walters 2023, Elkhatat 2023, Chaka 2023 and Ha and Passananti 2024 link the new cards; the source list of the LMS integrations slice, cut short in the file at launch, completed | the index data file and the vendor pages read 9 November 2026 |
| 2026-11-09 | all | Index version 2026.11.5, roadmap wave 4: 10 cards (AIDI-D-0029 to 0038: Smodin, BrandWell, Surfer, Compilatio, Quetext, TruthScan, DetectGPT, Sightengine, Deepware, Sensity); 2 rankings (platform integrations, 34 of 38 qualify, Walter Writes first at 6 of 8 surfaces; file types, 19 of 38, Sensity first at 22 formats), 4 questions (short texts, ChatGPT coverage, College Board, Spanish); rankings re-sorted: independently tested 14 to 17 of 38 (DetectGPT 3, BrandWell 2, Compilatio 1), false positive rate 12 to 15 (three zeros from Weber-Wulff 2023; Compilatio's vendor rate to the vendor table), languages 13 to 19 (Smodin 100+ first), input limits 8 to 15 (AI or Not first at 500,000 characters per API request, the third-wave cards recounted), data retention 9 to 13 (Sightengine and Smodin 0 days, TruthScan 30, Compilatio 1,825), studies by content type 16 to 20, adversarial robustness 8 to 9 (DetectGPT 7.6%), false negative rate 12 to 15, claim vs evidence gap 8 to 9 (Compilatio plus 15 points); Sensity's three-year biometric clause kept out of the retention field and TruthScan's image false positive figure out of the text field; the study tables of Weber-Wulff 2023, Walters 2023, Sadasivan 2023 and Krishna 2023 link the new cards; the API, LMS, modalities, open-source and published false positive rate slices and the business, agencies, recruiters and school districts views gained the qualifying cards, the API and modalities slices also gaining the third-wave cards they had missed; the ownership question records TruthScan and Undetectable AI as one company, Undetectable Inc. | the index data file, the vendor pages and the College Board pages read 9 November 2026 |
| 2026-11-09 | all | Index version 2026.11.6, roadmap wave 5: 4 cards (AIDI-D-0039 to 0042: Codeleaks, Illuminarty, Pindrop Pulse, Resemble Detect); 2 study records (AIDI-S-0018 Pegoraro 2023, patched onto nine cards, and AIDI-S-0019 Fleckenstein 2024, a human-rater study feeding no card; AIDI-S-0020 reserved for Dalalah 2023, held because no primary copy is reachable); 6 comparisons (AIDI-C-05 to 10: Winston AI vs Turnitin, GPTZero vs Grammarly, ZeroGPT vs Scribbr vs QuillBot, Pangram vs Originality.ai, Winston AI vs Originality.ai, Scribbr vs Turnitin); 2 views (AIDI-U-07 marketplaces and review platforms, 2 cards, and AIDI-U-08 researchers, 19 cards); 1 question (AIDI-Q-16 does Gradescope have AI detection?: no, by Turnitin's FAQ and Gradescope's feature table); rankings re-sorted: independently tested 17 to 18 of 42 (Pegoraro 2023 lifts GPTZero to 12, the GPT-2 Output Detector to 10, Copyleaks to 7, Originality.ai and ZeroGPT to 6, Writer to 5, GLTR and DetectGPT to 4, BrandWell to 3; Illuminarty joins at 1; Originality.ai and ZeroGPT move above Turnitin and BrandWell above Pangram on the counts), false positive rate 15 of 42 unchanged in membership with ZeroGPT above Turnitin on the tie rule, languages 19 to 21 (Resemble Detect 54, Pindrop Pulse 40), studies by content type 20 to 24 (Illuminarty 1; Codeleaks, Pindrop Pulse and Resemble Detect 0), false negative rate 15 to 16 (Illuminarty 4.69% on images), platform integrations 34 to 38 (Resemble Detect 3 of 8; three at 1 of 8; ZeroGPT above Turnitin on the tie rule), file types 19 to 20 (Resemble Detect 17), free tiers, input limits, data retention, adversarial robustness and claim vs evidence gap unchanged; four rulings: Illuminarty's 97.5% release-post validation figure is not paired as a claim, Illuminarty's purpose-bound clause and Pindrop's voice-features and Resemble Detect's flagged-media clauses state no retention period for submitted media (all three quoted in the second table), and Pegoraro's printed true positive and true negative rates feed no rank field because the false negative rule takes no complement; the Pegoraro record's sentence saying the complements feed the rankings was corrected to that rule; the Ha and Passananti 2024 table links the Illuminarty card; the code, modalities and API views, the shared owner question (Copyleaks and Codeleaks, one company) and the withdrawn detectors question (Copyleaks' AI code verdict removed from its API on 16 December 2025 per the release notes) gained rows; the For nav panel lists the two new views. | |
| 2026-11-09 | all | Index version 2026.11.7, roadmap wave 6: 11 study records (AIDI-S-0021 Ibrahim 2023, 0023 Kar 2024, 0024 Alshammari and Rao 2025, 0025 DAMAGE 2025, 0026 Epoch AI 2026, 0027 ASVspoof 5, 0028 SAFE 2025, 0029 DFDC 2020, 0030 RAID image 2025, 0031 Vanderbilt 2023 statement, 0032 GPTZero and Pangram vendor benchmarks; AIDI-S-0022 reserved for Jiang 2024, held because no primary copy is reachable); study counts GPTZero 12 to 16, GPT-2 Output Detector 10 to 12, Copyleaks 7 to 9, Originality.ai 6 to 7, ZeroGPT 6 to 7, Pangram 2 to 4, Sapling 2 to 3, Hive 1 to 2, QuillBot 0 to 2, Undetectable AI 0 to 1, Sightengine 0 to 1 (the Vanderbilt statement is a note on the Turnitin card and the vendor benchmarks are benchmark lines on the GPTZero and Pangram cards, counted in no field); three fields moved by rule: GPTZero robustness 20% to 60.04% (DAMAGE 2025, 5% FPR after 19 humanizers), Pangram false negative rate 1.31% to 0% (Epoch AI 2026, 0 of 297), GPT-2 Output Detector false negative rate 33.3% to 8.3% (Ibrahim 2023, 10 of 120); Hive and Sightengine gained image robustness values (44%, 0%, RAID image 2025); no false positive rate moved (Epoch AI's 0 of 495 is not lower than the zeros on the cards); 2 rank fields (minimum input, API price per 1,000 words) and 2 rankings (minimum input, 10 of 42, Phrasly first at 30 words; API pricing, 10 of 42, AI or Not lowest at $5 per million words); 2 updates (Q1 2027 index update, Q1 2027 studies digest); rankings re-sorted: independently tested 18 to 21 of 42, adversarial robustness 9 to 11, studies by content type 24 with Hive first at 2 records; register corrections: ADI-085 slug and title (Kar 2024), ADI-086 title (Alshammari and Rao 2025), ADI-087 venue (GenAIDetect at COLING 2025), ADI-091 metric (65.18% is accuracy in Meta's post, not average precision), ADI-092 Hive figures (0.60 clean, 0.44 adversarial, 1.00 real) | the index data file, arXiv 2501.03437, epoch.ai, DOI 10.1186/s40468-023-00260-2, arXiv 2506.03988, all read 9 November 2026 |
| 2026-11-09 | all | Index version 2026.11.8, roadmap wave 7: 11 question pages (AIDI-Q-17 minimum word count, Q-18 same result twice, Q-19 published test data, Q-20 offline and on-premise, Q-21 results as evidence, Q-22 text and images, Q-23 free API tiers, Q-24 paraphrased text, Q-25 tested on GPT-4, Q-26 false positive studies, Q-27 Canvas without Turnitin); 5 rank fields and 5 rankings (AIDI-R-15 LMS coverage, 11 of 42, Compilatio first at 7 of 9 platforms; R-16 free checks per day, 4 of 42, Sightengine first at 500 operations a day; R-17 generators covered, 36 of 42, Resemble Detect first at 62 named; R-18 non-English evidence, 14 of 42, GPTZero first with 5 records; R-19 code languages, 3 of 42, Codequiry first at 12); no study record, study count or evidence field changed; five corrections filed: every character-to-word conversion on the input limits ranking, the short texts question, the admissions view and the methodology table moved from 6 to 5.5 characters a word, one rate sitewide (Copyleaks 255 characters about 46 words, Smodin 1,000 about 182, AI or Not 500,000 about 91,000), with the Jabarian and Imas 2025 stub test under 50 words added to the short texts question; Compilatio's image cell on detectors by content type cleared to follow the card's modalities field (handwriting OCR is not image detection); the false positive evidence review gained DAMAGE 2025 (GPTZero 1.47%, human count not printed), Pegoraro 2023 true negative rates as printed and the Ha and Passananti 2024 image rows for AI or Not (24.47%) and Illuminarty (67.40%); 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 the false negative ranking cites; Turnitin's paraphrase (6 December 2023) and bypasser (27 August 2025) dates checked against the card and the claims record, consistent; the Originality.ai 5.0 post was read live on 9 November 2026 for the code languages field | the index data file and originality.ai, read 9 November 2026 |
| 2026-11-09 | all | Index version 2026.11.9, roadmap wave 8: 8 question pages (AIDI-Q-28 Google Docs, Q-29 public leaderboards, Q-30 bullet points, tables and code, Q-31 the Weber-Wulff 2023 roster, Q-32 voice and video together, Q-33 model cards, Q-34 Turnitin's languages by date, Q-35 deleting your data), 5 use-case views (AIDI-U-09 law firms, U-10 grant reviewers, U-11 newsrooms, U-12 localisation teams, U-13 government) and 3 feature slices (AIDI-F-06 no account, F-07 bulk and batch, F-08 compliance); four flags written into every card for the views (no_account, bulk, compliance, deployment, each with source and read date); the Walter Writes integrations field corrected from "Docs" to "Google Docs" after the extension page was re-read; the SynthID sources map gained the Nature DOI the card already cites; the Weber-Wulff 2023 record gained the four missing Table 5 rows (Check For AI 32, Crossplag 37, PlagiarismCheck 21, Writeful 23) and the Table 9 and 10 false accusation and undetected counts for all 14 tools from the preprint; the shut-down question records that 9 of the 14 Weber-Wulff tools now have a card, not 6; the GPTZero card sets its FAQ claim of 95.7% at 1% beside the RAID leaderboard entry of 7 October 2025 (97.24% at 5%, 92.93% at 1%); the Turnitin card gained the 4 April 2023 launch date and the Spanish (12 September 2024) and Japanese (2 April 2025) model dates; the count of cards quoting an accuracy figure now counts a percentage figure rather than any digit (30 of 42, previously 37) | the index data file, arXiv 2306.15666, github.com/liamdugan/raid, grants.nih.gov NOT-OD-23-149, nsf.gov and five vendor trust pages, all read 9 November 2026 |
| 2026-11-10 | all | Index version 2026.11.10, roadmap wave 9: 9 cards (AIDI-D-0043 to 0051: Decopy, Content Detector AI, GPTKit, PlagiarismCheck, Isgen, AI Detector Pro, Fast-DetectGPT, Paperpal, Detecting-AI.com); a tenth researched card, Imagga (AIDI-D-0052), held unpublished because the vendor publishes no AI-generated image detector on any live page, documentation section or archive capture read 10 November 2026; 2 comparisons (AIDI-C-11 Scribbr vs QuillBot, C-12 Reality Defender vs Sensity); 2 feature views (AIDI-F-09 model changelogs, 10 cards; F-10 institution-only, 5 cards); 2 updates (AIDI-UP-06 Q2 2027 index update, UP-07 Q2 2027 studies digest); no study record added; seven records gained a card row (Weber-Wulff 2023 for PlagiarismCheck, now 10 of 14 tools carded; Walters 2023, Kar 2024 and Alshammari and Rao 2025 for Content Detector AI; Orenstrakh 2023 and Perkins 2024 for GPTKit; RAID 2024 for Fast-DetectGPT); three data rulings: PlagiarismCheck and Isgen retention left unranked because each vendor publishes two conflicting statements (the Sensity and Pindrop reading), PlagiarismCheck LMS count 6 of 9 with Populi outside the nine-platform list, GPTKit claim gap 5.5 against the 87.50% mixed-set accuracy with the 18-point alternative noted; fifteen rankings gained members and Detecting-AI.com took first place in input limits at 1,000,000 characters per paid scan, with no existing order moved; the shut-down question gained the ContentDetector.ai redirect row from the Wayback CDX index and the shared-owner question the Content at Scale lineage note from the archived page; one quoted adverb in the Fast-DetectGPT paper's limitations sentence elided as "[...]" for the estate AI-slop gate | the index data file, the Wayback Machine CDX index for contentdetector.ai, arXiv 2306.15666, arXiv 2307.07411, arXiv 2405.07940 and the nine vendors' pages, read 9 and 10 November 2026 |
| 2026-11-10 | all | Index version 2026.11.11, roadmap wave 10: 12 cards (AIDI-D-0053 to 0060 and 0062 to 0065: Binoculars, Monica, Trinka, StealthWriter, Wordtune, Check For AI, Writefull, Passed.ai, ElevenLabs AI Speech Classifier, Intel FakeCatcher, Deepfake-O-Meter, Ircam Amplify); a thirteenth researched row, Optic (AIDI-D-0061), held as a duplicate of the AI or Not card, which already records the tool under its earlier name; 2 comparisons (AIDI-C-13 Pangram vs Copyleaks, C-14 QuillBot vs Turnitin); 1 view (AIDI-U-14 self-publishing authors, 12 cards, the KDP Content Guidelines fetched live); no study record added; five records gained a card row (Weber-Wulff 2023 for Check For AI and Writefull, now 12 of 14 tools carded; Pegoraro 2023 for Writefull; Kar 2024 for Wordtune; RAID 2024 and DAMAGE 2025 for Binoculars); data rulings: Binoculars bulk flag off (a README list argument is not a product batch feature), Wordtune free checks per day and language count left unranked (bundle-only values of a withdrawn tool, quoted), Writefull retention left unranked (an archived statement for a discontinued tool), Deepfake-O-Meter retention left unranked (an API tier value with no policy page), and Writer's generator count recounted from 2 to 1 so that ChatGPT and GPT-4 count as one GPT family on every card; sixteen rankings gained members with no existing order moved except Writer's recount; the shut-down question gained four rows (Wordtune, Check For AI, Writefull, Passed.ai) and the Weber-Wulff roster three card links; three vendor figures written with the accuracy word bracketed for the claim gate (ElevenLabs, FakeCatcher, Ircam Amplify) | the index data file, the Wayback Machine CDX index for wordtune.com, checkforai.com, x.writefull.com and passed.ai, arXiv 2306.15666, arXiv 2304.01487, arXiv 2405.07940, arXiv 2501.03437, kdp.amazon.com and the twelve vendors' pages, read 10 November 2026 |
| 2026-11-10 | all | Index version 2026.11.12, roadmap wave 11: 9 cards (AIDI-D-0066 to 0074: GPT Radar, Grammica, IvyPanda, SEO.ai, AI Writing Check, RADAR, Ghostbuster, Hiya Deepfake Voice Detector, McAfee Deepfake Detector); 5 study records (AIDI-S-0033 to 0037: Rashidi 2023, ASVspoof 2021, ADD 2023, FaceForensics++ 2019, the Originality.ai compilation, a vendor statement on the Originality.ai card's vendor benchmarks line); 1 question (AIDI-Q-36 sentence-level highlighting); 1 ranking (AIDI-R-20 launch year); the GPT-2 Output Detector gains Rashidi 2023 as a 13th record and its false positive base gains the up to 8.69% of 1,600 figure while the field keeps 0 of 10; four records gained a card row (Walters 2023 for GPT Radar, Grammica, IvyPanda and SEO.ai, now 14 of 16 tools carded; Liang 2023 for AI Writing Check; RAID 2024 and DAMAGE 2025 for RADAR; NIST AI 700-1 names RADAR as a note); data rulings: RADAR's NIST naming is a note and not a record (2 records), its Space endpoint is not a product API, GPT Radar's 400-word advice and Ghostbuster's 100-word refusal are quoted and not ranked as minimums, Grammica and AI Writing Check are withdrawn and carry no no-account flag, McAfee's 2024 launch price leaves the plan table and stays in prose, "on-device (Windows AI PC NPU)" added to the deployment vocabulary; register corrections: the pathology abstracts study is Rashidi et al. 2023 (J Pathol Inform 14, 100342, DOI 10.1016/j.jpi.2023.100342), not the PMC id first noted, and the Originality.ai post lists 17 studies, not 14 | the index data file, the primary papers and the vendor pages named on each page |
| 2026-11-10 | all | Index version 2026.11.13, roadmap wave 12, the final wave: no card or record added; 4 questions (AIDI-Q-37 student plans, Q-38 scanned PDFs, Q-39 music detectors, Q-40 nonprofit and education pricing); 2 comparisons (AIDI-C-15 Sapling vs GPTZero, C-16 Pindrop Pulse vs Resemble Detect); 4 views (AIDI-U-15 customer support QA, AIDI-F-11 team seats and SSO, F-12 handwriting and OCR, F-13 mobile apps); 3 updates (AIDI-UP-08 the 2027 census, UP-09 the Q3 2027 index update, UP-10 the Q3 2027 studies digest); a team field (seats and SSO flags) written to every card, 20 of 72 set; the Turnitin card gains the capabilities FAQ's typed-text sentence from the 10 November 2026 re-read; two field values corrected under the file types ranking's stated token rule, Resemble Detect 17 to 19 upload formats (.amr and .3gp) and Ircam Amplify 7 to 9 (AIF and AIFF), the counting vocabulary of 9 November 2026 having dropped the four tokens; Ircam Amplify moves from 13th to 12th in that ranking, the first order change made by a correction; the Deepware Google Play listing re-read live (404); the roadmap's four held rows (Imagga, Optic, Dalalah 2023, Jiang 2024) listed in the census for a decision | the index data file, the Turnitin capabilities FAQ (article 28477544839821, edited 3 November 2026), docs.resemble.ai/detect/create, docs.ircamamplify.io audio input methods, play.google.com |
Each line of the log is written so that it can be read without the page it refers to. The date is the day the change went live, in the same format as every "Updated" line on the site. The page column names the card, record, ranking or view by its title and links it. The change column states the field and, for a figure, the old value and the new value in the same sentence. The source column links the vendor page or the paper that supports the new value and gives the date it was read, so that a reader can check the correction the way the editors did.
How a correction is filed
Send the page address, the field and the URL of your source to editor@aidetectorindex.com; the editors read the source, change the page the day the correction is confirmed, and add a line to the table above with the old and new value. The contact page sets out the same three items.
Planned rechecks
Recheck plans are recorded here with a date rather than stated as a cadence. Prices and free tiers on all 72 cards are planned for recheck on 8 December 2026, and each change that recheck produces will be logged as its own line with the page, the field, the old value and the new value.
How a logged change moves through the site
Every figure lives once, in the index data file, and each card, ranking and page schema reads from it at build time. A corrected price therefore changes on the card, in the free tiers ranking and in the page markup in the same build, and a corrected study result can move a row on the independently tested ranking the same day. The methodology page describes that rule and the 14 fields a correction can touch.
The editorial policy sets the rules a correction is judged against; the methodology page explains the fields and rankings; the contact page is the route for filing a correction; and the about page describes the desk that keeps this log.