AI detectors that publish model changelogs: dated release notes, model cards and version posts
An AI detector vendor publishes a dated record of model changes for 10 of the 72 cards in the index, by the vendor pages recorded on each card and read on 8, 9 and 10 November 2026: a release-notes page (Turnitin, Copyleaks and its Codeleaks product, Sightengine, Sapling), a model card with a date (Pangram, two cards dated 29 July and 19 October 2026), or a dated post that names a model version and what changed (GPTZero's 4o post of 18 October 2026, Originality.ai's 5.0 posts and accuracy blog of 25 October 2026, PlagiarismCheck's benchmark post of July 2026, Compilatio's what's-new entry of 5 November 2026 for detector v4.6.2). The latest dated entry the index read is Compilatio's, 5 November 2026, and the oldest still current is Copyleaks' release note of 16 December 2025 removing the AI verdict for source code (compilatio.net, docs.copyleaks.com, read 9 November 2026). A changelog is a vendor statement about its own model and not a test of it; the table records where each was read and when, and the independently tested ranking is where the outside results sit. This page is one of the views under AI detectors by feature; the model card question covers the narrower question of a published model card, and the full index shows all 72 cards.
- Field 1Changelog, release notes or model card
- Field 2Latest dated entry read
- Field 3What the entry records
- Field 4Read date
No single rank field holds this; the view is compiled from the source pages recorded on each card. A card qualifies when a vendor page already cited on the card carries a dated entry that names a model, a version or a model change: a release-notes or changelog page with dated entries, a model card that states its date, or a vendor post that names a model version and a release date. A pricing page that names a model version with no date (Pangram 4.1 on the API price line), a product label with no history (GPTKit's "v3", Detecting-AI.com's "detector v3") and a paper's own arXiv version history (Fast-DetectGPT) do not qualify on their own and sit in the second table. The "latest dated entry" column is the newest date the index read on that page, not a claim that the page has no newer entry: the pages were read on the dates shown and are not monitored. Field rules are in the methodology.
Which field decides whether a detector publishes a changelog?
| Detector | Changelog, release notes or model card | Latest dated entry read | What the entry records | Read date |
|---|---|---|---|---|
 Compilatio | compilatio.net what's new | 5 November 2026 | Detector v4.6.2, live for Magister+ and Studium since 1 October 2026, trained on "authentic human texts, compared with texts generated by the 2026 versions of ChatGPT, Claude, and Gemini in several languages" | 9 November 2026 |
 Originality.ai | 5.0 and 5.0 Max launch post; accuracy blog | 25 October 2026 (both posts) | 5.0 and 5.0 Max: AI recall 97.69% and 98.00% on the UChicago Booth set, 99.33% on Epoch AI; humanizer detection 89.52% and 92.38% on 420 humanized outputs with no human controls; the accuracy blog states "99%+ accuracy" for both | 8 November 2026 |
 Pangram | Pangram 4 model card; Pangram 4.1 model card | 19 October 2026 (4.1) | The 4.1 card reports 70 false positives in 781,981 human FineWeb texts (0.0090%); the 4 card is dated 29 July 2026; the vendor benchmarks record holds both as vendor measurements | 9 November 2026 |
 GPTZero | GPTZero 4o launch post; benchmarking page | 18 October 2026 (4o post) | 4o "deliberately tackles the harder cases for AI detection: AI-generated text that has been paraphrased or mixed with human writing", with "a false positive rate of only 1 in 10,402" on student writing; the benchmarking page (5 February 2026, updated 20 July 2026) gives 0.08% for model 4.3b | 8 and 9 November 2026 |
 Sapling | Detector API docs, changelog entry | 17 September 2026 | "Significantly higher detection of humanized AI text", up 3.9% absolute, and human-text false positives down 60% on an internal held-out modern-text set with the model upgrade | 8 November 2026 |
 Copyleaks | docs.copyleaks.com release notes; methodology page (model V11, test date 10 September 2026, published 28 September 2026) | 28 September 2026 (methodology); 16 December 2025 (release notes entry the index read) | The methodology page reports model V11 at 923 of 326,120 human texts flagged (0.9970 human-only accuracy); the release notes record the AI verdict for source code deprecated on 28 August 2025 and removed from the submit endpoints on 16 December 2025 | 8 and 9 November 2026 |
 Codeleaks | docs.copyleaks.com release notes, the same page as Copyleaks | 16 December 2025 | "The properties.aiGeneratedText" removed for source code: from that date the AI detection feature "will no longer support source code files"; plagiarism and licence scanning of code continue | 9 November 2026 |
 PlagiarismCheck | Benchmark post, July 2026 | July 2026 | Seven models tested (Gemini 3.1 Pro, Gemini Flash 3.5, Claude Sonnet 4.6, Claude Opus 4.7, GPT-5.4, GPT-5.5, Grok 4.3): 99.23% of 33,699 AI sentences classified as AI, 1.04% of 280,781 human sentences flagged, 0.00% of 2,805 human documents at the over-50% threshold; the detector page calls the figures marketing | 9 November 2026 |
 Sightengine | sightengine.com/docs/changelog | June 2026 | The genai image model's changelog begins in March 2025; a speech model arrived in November 2025 and a music model in June 2026, per the dated entries | 9 November 2026 |
 Turnitin | AI writing detection model release notes | 5 May 2026 (Spanish model update) | Launch 4 April 2023; paraphraser detection 6 December 2023; Spanish model 12 September 2024, updated 5 May 2026; Japanese model 2 April 2025; bypasser detection 27 August 2025; the FAQ states that before every model release Turnitin tests "over 700,000 additional academic papers" written before ChatGPT | 8 and 9 November 2026 |
Entries as each vendor dates them, read 8, 9 and 10 November 2026 and ordered by the latest dated entry; a row names the newest entry the index read, and the page may carry later entries since. Every figure in the "what the entry records" column is the vendor's own measurement of its own model, which is why none of them enters a rank field; the claim vs evidence gap ranking is where a vendor figure meets a study figure.
Labels as the cards record them, read 8 to 10 November 2026; the leaderboard question covers the RAID entries.
Why each card is in this view
- AIDI-D-0032Compilatio
the newest dated entry in the index, a what's-new post of 5 November 2026 for detector v4.6.2 that names its training comparison set (compilatio.net, read 9 November 2026); the one study record, Weber-Wulff 2023, tested the free tool in March 2023, three years and several versions earlier.
- AIDI-D-0004Originality.ai
two dated posts that name 5.0 and 5.0 Max and print the vendor's recall and humanizer figures (originality.ai/blog, read 8 November 2026); the card carries 7 study records, the most recent Epoch AI 2026 on Turbo 3.0.2, a version the vendor's posts have since superseded.
- AIDI-D-0006Pangram
the only vendor in the index that publishes a model card per release with a date and a false positive count on a named corpus (pangram.com/research/model-card/pangram-4-1, read 9 November 2026); the cards are recorded in the vendor benchmarks record and feed no field.
- AIDI-D-0001GPTZero
a launch post per model (4o on 18 October 2026) and a benchmarking page that dates its own updates (gptzero.me/news, read 8 November 2026); the 16 study records on the card span models from 2023 to 2026-05-11-base, so a reader can match each result to the model generation it tested.
- AIDI-D-0010Sapling
a changelog entry on the API documentation page dated 17 September 2026 with two internal figures (sapling.ai/docs/api/detector, read 8 November 2026); the three study records predate the entry.
- AIDI-D-0003Copyleaks
a release-notes page and a dated methodology page naming model V11 (docs.copyleaks.com, read 9 November 2026); the release note of 16 December 2025 is the index's example of a changelog recording a withdrawal rather than an improvement.
- AIDI-D-0039Codeleaks
the same release notes, which are where the vendor says the AI verdict for code no longer exists while the product page still describes code detection (Codeleaks card, read 9 November 2026).
- AIDI-D-0046PlagiarismCheck
a benchmark post dated July 2026 naming seven models and the sentence and document counts behind the 99.23% (plagiarismcheck.org/blog, read 9 November 2026); the one study record, Weber-Wulff 2023, scored 39% of 54 documents in May 2023, the widest gap in the claim vs evidence gap ranking.
- AIDI-D-0036Sightengine
a documentation changelog with dated model arrivals from March 2025 to June 2026 (sightengine.com/docs/changelog, read 9 November 2026); the one record, RAID image 2025, prints no API version or date, so the changelog cannot be matched to it.
- AIDI-D-0002Turnitin
a model release-notes page from the 4 April 2023 launch to the 5 May 2026 Spanish update (guides.turnitin.com, read 9 November 2026); the five records on the card ran the detector between May 2023 and 2026 and the notes let a reader place each between releases.
What the evidence says about model changes
A changelog dates a vendor's model, and a study dates a test; the index can match the two only where both print a date. Turnitin's notes place the paraphraser detection of 6 December 2023 after Weber-Wulff 2023's May 2023 run, in which Turnitin missed 10 of 36 AI documents, all among the edited or paraphrased texts, and before Van Vlasselaer 2026, in which all 40 fully AI papers from ChatGPT Deep Research scored under 20% (Weber-Wulff 2023; Van Vlasselaer 2026). GPTZero's records name the model where the study did: Epoch AI 2026 tested model 2026-05-11-base and DAMAGE 2025 an earlier model, while the 4o post of 18 October 2026 describes a model no indexed study has run (Epoch AI 2026; DAMAGE 2025). Originality.ai's 7 records end at Turbo 3.0.2 in Epoch AI 2026, and the vendor's 5.0 posts report figures for a model the index has no outside result on. Compilatio's v4.6.2 note and PlagiarismCheck's July 2026 benchmark both postdate every record on their cards by more than three years.
The pattern across the ten is that vendors publish changelogs as the evidence of improvement, and the index records them as dated claims. Six of the ten entries carry a vendor accuracy or false positive figure for the new model (Originality.ai, Pangram, GPTZero, Sapling, Copyleaks, PlagiarismCheck), and none of those figures enters a rank field, because each was measured by the vendor on a set the vendor chose; the false positive rate ranking sorts only study-sourced rates and lists the vendor figures in a second table. The other value of a changelog is negative evidence: Copyleaks' release notes are the only place the index found the statement that the AI verdict for code was withdrawn, a fact the product page does not carry (shut-down question). A reader who wants to know whether a study result still describes the tool should read the changelog date against the study's test date, and the Q2 2027 index update lists the dated changes the index itself recorded in the period. No detector score should be the only evidence in a misconduct, hiring or publishing decision, and a changelog entry that reports a better score is a reason to look for a new outside test, not a replacement for one.
Sources
- https://guides.turnitin.com/hc/en-us/articles/28294949544717-AI-writing-detection-model (read 9 November 2026)
- https://www.pangram.com/research/model-card/pangram-4-1 (read 9 November 2026)
- https://www.pangram.com/research/model-card/pangram-4 (read 9 November 2026)
- https://gptzero.me/news/introducing-gptzero-4o/ (read 8 November 2026)
- https://gptzero.me/news/gptzero-ai-detection-benchmarking-the-industry-standard-in-accuracy-transparency-and-fairness/ (read 9 November 2026)
- https://docs.copyleaks.com/resources/updates/release-notes (read 9 November 2026)
- https://sightengine.com/docs/changelog (read 9 November 2026)
- https://originality.ai/blog/introducing-originality-5-0-and-max (read 8 November 2026)
- https://originality.ai/blog/ai-accuracy (read 8 November 2026)
- https://plagiarismcheck.org/blog/ai-detection-accuracy-and-method-evidence-numbers-and-competitor-comparison/ (read 9 November 2026)
- https://sapling.ai/docs/api/detector/ (read 8 November 2026)
- https://www.compilatio.net/en/whats-new/introduction-to-the-compilatio-ai-detector-v4-6-2/ (read 9 November 2026)
- https://huggingface.co/openai-community/roberta-base-openai-detector (read 9 November 2026)
- https://github.com/baoguangsheng/fast-detect-gpt (read 10 November 2026)
Revision log: 2026-11-10: view created at index version 2026.11.10
FAQ: AI detector model changelogs
Which AI detectors publish a model changelog or release notes?
Ten cards link a dated record on a vendor page: Turnitin (model release notes), Copyleaks and Codeleaks (release notes), Sightengine and Sapling (changelog entries), Pangram (dated model cards), GPTZero, Originality.ai and Compilatio (dated version posts) and PlagiarismCheck (a dated benchmark post), read 8 to 10 November 2026.
Which detector has the most recent dated model entry?
Compilatio, whose what's-new post of 5 November 2026 describes detector v4.6.2 (compilatio.net, read 9 November 2026); Originality.ai's accuracy blog of 25 October 2026 and Pangram's 4.1 model card of 19 October 2026 follow.
How often do AI detectors update their models?
The dated entries the index read give two to four dated changes a year for the vendors that publish them: Turnitin lists six dated model events from April 2023 to May 2026, Pangram two model cards in 2026 (29 July and 19 October), GPTZero two dated pages in 2026, and Sightengine three model arrivals between March 2025 and June 2026. Vendors with no changelog publish no cadence the index can quote.
Does a changelog figure count as evidence in the index?
No. The figures in release notes and model cards are the vendor's own measurements and are recorded as such, in the vendor benchmarks record for GPTZero and Pangram and on the cards for the others; only a study in the studies index feeds a rank field.
Which cards state a version with no changelog?
GPTKit ("v3") and Detecting-AI.com ("detector v3") label a version with no release history; the GPT-2 Output Detector has a model card with no dated changes; Fast-DetectGPT has a paper version history and a repository push date; Winston AI and Grammarly have dated RAID leaderboard entries but no vendor changelog on the pages read.
