Which AI detectors publish a model card or technical report? The eight cards with a dated, sized document, and the 34 with a claim instead

One, in the strict sense: Pangram, whose model cards for Pangram 4 and Pangram 4.1 carry a version, a date, named evaluation sets with sizes and a stated calibration target, "roughly 1 in 10,000" false positives (pangram.com, read 9 November 2026). The 4.1 card changed the headline rate from 41 in 1,000,000 to 70 in 781,981 between July and October 2026 because the FineWeb test set was cleaned of "spam, advertising and programmatically generated documents", which the vendor benchmarks record notes is a change of test set and not only of model. GPTZero's benchmarking page does the same work under another name: it states the model, 4.3b, the construction of four English sets of 1,000 human and 1,000 AI texts, the generators used and the binarisation rule, and it says three of the four sets cannot be published (gptzero.me, read 9 November 2026). Copyleaks' methodology page states model V11, a test date of 10 September 2026 and 326,120 human texts, of which 923 were flagged (copyleaks.com, read 8 November 2026). Resemble Detect's benchmarks page states file counts for audio, image and video and a third-party audio benchmark, Podonos, with competitors' figures beside its own (resemble.ai, read 9 November 2026). A fifth company, OpenAI, published a dated account of a detector it no longer runs: the 31 January 2023 announcement of its AI Text Classifier reported 26% of AI text caught and 9% of human text flagged on an unsized English challenge set, and the same page carries the withdrawal of 20 July 2023 (openai.com, read 8 November 2026; OpenAI classifier record). It has no card because the tool is gone, so it sits in this paragraph and not in the table.

The wider question, which vendors publish any dated record of model changes, is answered on the model changelogs view, added 10 November 2026: 10 cards, from Turnitin's model release notes and Copyleaks' release notes to Compilatio's what's-new entry of 5 November 2026, with the latest dated entry read on each page.

Four, and in three of them the paper is the product. DetectGPT is a method from a January 2023 paper by Mitchell, Lee, Khazatsky, Manning and Finn at Stanford, which reports AUROC on 500 samples per dataset for five open models and publishes no accuracy percentage or false positive rate, because AUROC is threshold-free (arXiv 2301.11305, read 9 November 2026). GLTR is the 2019 demo paper by Gehrmann, Strobelt and Rush, whose one quantified claim is that its colouring raised human detection of fake text from 54% to 72% in a human-subjects study (arXiv 1906.04043, read 9 November 2026). The GPT-2 Output Detector is the RoBERTa classifier OpenAI released with the 1.5B GPT-2 model on 5 November 2019, with "detection rates of ~95%" in the release post and a Hugging Face model card that gives the training and evaluation sets (openai.com; huggingface.co, read 9 November 2026). SynthID is different in kind: the Nature paper by Dathathri and colleagues at Google DeepMind describes a watermark embedded at generation and measures how detectable it is at a fixed 1% false positive rate on the authors' own outputs, so it reports on text the vendor's models wrote, not on text found in the wild (DOI 10.1038/s41586-024-08025-4, read 9 November 2026). The open-source view lists the three of these four whose code is public.

A number without a base, a base without a date, or no number. By the vendor claim field of each card read 9 November 2026 in the data file, 17 of the 34 state a numeric accuracy or detection rate of their own and 17 state none, quoting a competitor's benchmark, a disclaimer or a superlative instead. Several name a dataset without its size on the pages read: Originality.ai's 5.0 launch post of 25 October 2026 gives recall on the "UChicago Booth AI" and "Epoch AI" sets (originality.ai, read 8 November 2026), Winston AI's home page states "zero false positives across 27,743 human documents in published evaluations" (gowinston.ai, read 8 November 2026), Proofademic's research page, published May 2026, states an accuracy range of 97.6% to 99.9% "across the eight evaluated model families" (proofademic.ai, read 8 November 2026), and Illuminarty's release post of 18 March 2023 gives validation accuracy "within our collected dataset" (illuminarty.ai, read 9 November 2026). Three cards cite RAID, a public benchmark, for their figure (QuillBot, Grammarly, Winston AI), which the leaderboard question covers from the leaderboard side. Turnitin publishes a false positive rate on 700,000 to 800,000 pre-ChatGPT papers and no accuracy figure, which the Turnitin claims record sets out; it is listed below because no document names the model version or a test date.

Cards with no model card or sized technical report

Claims as the vendor claim field of each card quotes them from the page linked, read on the date shown; 8 + 34 = 72, counted 9 November 2026 in the data file.

A model card tells a reader what the vendor measured and on what, and the eight that exist were all written by the people selling or publishing the detector.

All read 9 November 2026 unless an earlier date is shown; the source of every row is linked in the tables.

• Pangram 4.1 and Pangram 4 model cards: https://www.pangram.com/research/model-card/pangram-4-1 and https://www.pangram.com/research/model-card/pangram-4
• Emi and Spero 2024, Technical Report on the Pangram AI-Generated Text Classifier: https://arxiv.org/abs/2402.14873
• GPTZero AI Detection Benchmarking and technology page: https://gptzero.me/news/gptzero-ai-detection-benchmarking-the-industry-standard-in-accuracy-transparency-and-fairness/ and https://gptzero.me/technology
• Copyleaks AI Detector testing methodology: https://copyleaks.com/ai-detector/testing-methodology (read 8 November 2026)
• Resemble AI benchmarks: https://www.resemble.ai/benchmarks
• Dathathri et al. 2024, Nature 634, 818 to 823: https://doi.org/10.1038/s41586-024-08025-4
• Mitchell et al. 2023, DetectGPT: https://arxiv.org/abs/2301.11305
• Gehrmann, Strobelt and Rush 2019, GLTR: https://arxiv.org/abs/1906.04043
• RoBERTa base OpenAI detector model card: https://huggingface.co/openai-community/roberta-base-openai-detector
• OpenAI, GPT-2 1.5B release: https://openai.com/index/gpt-2-1-5b-release/
• OpenAI, New AI classifier for indicating AI-written text: https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/ (read 8 November 2026)
• Turnitin AI writing detection FAQs: https://guides.turnitin.com/hc/en-us/articles/28477544839821-Turnitin-s-AI-writing-detection-capabilities-FAQs (read 8 November 2026)
• Originality.ai, Introducing Originality 5.0 and Max: https://originality.ai/blog/introducing-originality-5-0-and-max (read 8 November 2026)
• Winston AI home page: https://gowinston.ai/ (read 8 November 2026)
• Proofademic AI text detector: https://proofademic.ai/ai-text-detector/ (read 8 November 2026)
• Illuminarty release post, 18 March 2023: https://illuminarty.ai/en/posts/big-spring-update.html
• Every other row's source is the vendor claim source recorded on its card, linked in the second table

Revision log: 2026-11-09: page created from the vendor claim fields of 42 cards and the vendor benchmarks record at index version 2026.11.8; 2026-11-10: model changelogs view linked