Which AI detector vendors publish their test data? Every benchmark page, model card and methodology article in the index, with what can be downloaded
Twelve of the 72 AI detector vendors in the index publish a dated evaluation of their own tool that names or sizes the test set, by the vendor claim and benchmark fields of every card read 8 and 9 November 2026 in the data file, and one of the twelve, GPTZero, publishes raw predictions for one of its four test sets while stating that the other three "cannot be published" for copyright reasons (gptzero.me, read 9 November 2026). The largest stated set is Pangram's 781,981 human FineWeb texts on its 4.1 model card of 19 October 2026, a public corpus whose scored subset is not published (pangram.com, read 9 November 2026); the one fully downloadable dataset on the record is not a vendor's at all but RAID, the shared benchmark of 6,287,820 texts with data and code on GitHub and a live leaderboard (arXiv 2405.07940, read 8 November 2026). 15 further vendors print a numeric claim with no test set behind it, and the remaining 15 cards print no accuracy figure for their own tool at all.
A published test is three things at once: a figure, the model or date it belongs to, and the set of texts or files it was measured on. The first table lists every card whose vendor states all three on a page the index read, with a fourth column for whether the data behind the figure can be obtained, as the card or the study record describes it. The second table lists the cards whose vendor prints a figure without a test set. RAID appears at the foot of the first table as the reference point: a benchmark built by researchers, not a vendor, whose data is on GitHub. This page is one of the questions answered from the index; its siblings ask which detectors claim a false positive rate under 1% and whether an AI detector is itself AI, the claim vs evidence gap ranking sets each vendor figure against the studies that tested the same tool, and the AI detector register carries every card.
| Detector | Published evaluation | Test set stated | Data obtainable, as recorded | Source | Read date |
|---|---|---|---|---|---|
| GPTZero | Benchmarking page, model 4.3b, 5 February 2026, updated 20 July 2026: false positive rate 0.08%, recall 99.60% | 1,000 human and 1,000 AI texts in each of four English domains; 1,000 bypassed AI texts | Raw predictions published for the academic paper review set; three English sets restricted to researchers on request | gptzero.me | 9 November 2026 |
| Pangram | Model cards for Pangram 4 (29 July 2026) and 4.1 (19 October 2026): 70 false positives in 781,981 texts (0.0090%) | 781,981 pre-ChatGPT English FineWeb texts; 520,000 AI generations from 26 models | FineWeb and FineWeb2 are named public corpora; the scored subsets are not published | pangram.com | 9 November 2026 |
| Copyleaks | Methodology page, model V11, tested 10 September 2026, published 28 September 2026: 923 of 326,120 human texts flagged, 0.30% false positives at the Balanced setting | 326,120 human texts; sensitivity table at three settings | No dataset link recorded on the card | copyleaks.com | 8 November 2026 |
| Originality.ai | 5.0 and 5.0 Max launch post and accuracy post of 25 October 2026: human specificity 99.954% on PERSUADE 2.0, 98.901% on the Liang TOEFL essays, 99.940% on PELIC | Named corpora: PERSUADE 2.0, the 91 Liang 2023 TOEFL essays, PELIC, UChicago Booth, Epoch AI | Corpora named; no scored file linked on the card | originality.ai | 8 November 2026 |
| Turnitin | FAQ and blog of 23 May 2023: document false positive rate under 1% for documents with over 20% AI writing | Over 700,000 pre-ChatGPT academic papers per release (800,000 in May 2023); composition not published | Not published; no third-party audit | guides.turnitin.com | 8 November 2026 |
| Compilatio | Support article: reliability between 94% and 99% and a false positive rate below 1% | About 7,400 texts in 24 languages, 3,700 human and 3,700 AI | Not published on the article read | support.compilatio.net | 9 November 2026 |
| Proofademic | Detector and research pages, evaluation of March 2026, study published May 2026: 0.2% document false positive rate, 97.6% to 99.9% (accuracy) across eight model families | 4,200 fully human-authored academic texts, held out | Not published on the pages read | proofademic.ai | 8 November 2026 |
| Winston AI | Home page: zero false positives across 27,743 human documents "in published evaluations"; 25,996 PERSUADE 2.0 essays, zero false positives | 27,743 human documents, 25,996 of them PERSUADE 2.0 essays | PERSUADE 2.0 named; the evaluations are not linked on the page read | gowinston.ai | 8 November 2026 |
| Resemble Detect | Benchmarks page, last updated August 2026: 99.5% pooled detection (accuracy) on the Podonos Audio DFD Bench; image false positive rate 4%, video 0.95% | 4,524 audio files; 1,000 image files; 1,000 video files; figures from the Hugging Face Speech DF Arena and DFBench leaderboards of March 2026 | Public leaderboards named; the vendor states it "does not control test set composition" | resemble.ai | 9 November 2026 |
| Pindrop Pulse | Audio deepfake detection page: "99% Detection accuracy against known engines", footnoted to the ASVspoof5 dataset of 32 engines; "<1% false positive rate" footnoted "Pindrop Product Data, 2024" | ASVspoof 5, 32 attack engines; a 20 million sample proprietary dataset for the meetings claim | ASVspoof 5 is an indexed challenge dataset whose papers print no Pindrop result; the proprietary set is not published | pindrop.com | 9 November 2026 |
| GPT-2 Output Detector | OpenAI release post: detection rates of about 95% for 1.5B GPT-2 text; model card: 510-token examples | 5,000 WebText and 5,000 GPT-2 samples per the model card | The samples live in the openai/gpt-2-output-dataset repository with the MIT-licensed weights | huggingface.co | 9 November 2026 |
| DetectGPT | Paper, arXiv 2301.11305: AUROC 0.97 on XSum, 0.92 on SQuAD, 0.97 on WritingPrompts across five open models | 500 samples per dataset from three named public corpora | Code under the MIT License on GitHub; the corpora are public | arxiv.org | 9 November 2026 |
| RAID benchmark (reference, not a vendor) | Dugan et al. 2024, ACL: 12 detectors on over 6 million texts | 6,287,820 texts: 11 generators, 8 domains, 11 attacks, 14,971 human documents | Data and code on GitHub; a live leaderboard at raid-bench.xyz; hidden test labels kept private | arXiv 2405.07940 | 8 November 2026 |
Figures as each vendor prints them on the page linked, read on the date shown; every vendor figure is measured by the vendor on a set the vendor chose, and none enters the false positive rate ranking, which takes study results only.
Which vendors publish a benchmark page or a model card?
Two publish a document built for the purpose. GPTZero's benchmarking page states the model (4.3b), the construction of each 1,000-and-1,000 set, the generators used (GPT-5.2, Gemini 3 Pro, Claude Sonnet 4.5 and Grok 4 Fast), the binarisation rule and the figures, and scores four competitors on the same sets (gptzero.me, read 9 November 2026). Pangram's model cards state the version, the set, the cleaning applied between the 4 and 4.1 cards, a 95% confidence interval of 0.0071% to 0.0113% on the 4.1 false positive rate, and reruns on third-party data (pangram.com, read 9 November 2026). Both sit in the vendor benchmarks record, which also notes that each vendor ranks itself first. Copyleaks publishes a methodology page with a test date, a model number and a base of 326,120 human texts (copyleaks.com, read 8 November 2026), and Resemble a benchmarks page that names outside leaderboards and dates them March 2026 (resemble.ai, read 9 November 2026). The other eight rows in the table state a sized set on a product page, a support article, a launch post or a paper rather than a page kept for the evaluation.
Can the test data itself be downloaded?
For one vendor set in part, for one research dataset in full, and for the rest in name only. GPTZero releases raw predictions for its academic paper review set and keeps the other three English sets for researchers on request, because "data usage policies and copyright constraints" prevent publication (gptzero.me, read 9 November 2026). Pangram, Originality.ai, Winston AI and Resemble name public corpora or leaderboards, FineWeb, PERSUADE 2.0, the Liang 2023 TOEFL essays, the Hugging Face Speech DF Arena, so a reader can obtain the corpus but not the vendor's scored subset or its predictions (pangram.com; originality.ai; gowinston.ai; resemble.ai, read 8 and 9 November 2026). Turnitin's 700,000 to 800,000 papers, Compilatio's 7,400 texts and Proofademic's 4,200 texts are counts with no published composition or file (Turnitin claims record). RAID is the exception the index holds: Dugan et al. 2024 released the data and code on GitHub and run a leaderboard, keeping only the hidden test labels private, which is why the RAID record quotes leaderboard values only with the date read (arXiv 2405.07940, read 8 November 2026). The two open-source detectors in the table, the GPT-2 Output Detector and DetectGPT, publish code and point at public samples, as the open-source view records.
A public leaderboard is the other place a vendor's test result can be read: 6 cards have an entry on RAID's, read 9 November 2026, and none appears under its own name in the NIST, ASVspoof 5, SAFE 2025 or DFDC tables, as the leaderboard question sets out. The model card question, added the same day, applies a narrower test, a document naming the model, the test set and its size, and finds 8 cards.
Which vendors print a figure with no test set behind it?
Fifteen, by the vendor claim field of each card read 8 and 9 November 2026.
| Detector | Figure printed | What stands behind it on the page read | Source | Read date |
|---|---|---|---|---|
| QuillBot | "99% detection rate, according to independent evaluations from RAID" | A reference to the RAID leaderboard; no QuillBot test set or date | quillbot.com | 8 November 2026 |
| Grammarly | "99% detection accuracy and ranks #1 on RAID's independent benchmark" | A reference to the RAID leaderboard; no Grammarly test set or date | grammarly.com | 8 November 2026 |
| Phrasly | 99.8% (accuracy) on fully AI-generated content | No set, date or model | phrasly.ai | 8 November 2026 |
| Sapling | False positives fell by 60% in the 17 September 2026 model upgrade | "Internal held-out evaluations"; a relative change, no absolute rate or set size | sapling.ai | 8 November 2026 |
| JustDone | 80% overall detection accuracy, 10.3% error rate, labelled "Independently benchmarked"; blog updated 19 March 2026 | No set size or tester named | justdone.com | 8 November 2026 |
| ZeroGPT | 98.4% detection (accuracy); under 1% false positive rate | "Internal evaluations" in the FAQ; no set or date | zerogpt.com | 8 November 2026 |
| Codequiry | Model v3e catches 90% of AI code and falsely flags 1.3% of human code; per-tool rates from 80% to 91% | A model name; no sample size published | codequiry.com | 9 November 2026 |
| Undetectable AI | "99%+ Accuracy Proven By Independent Tests"; "100% detection / 0% false positives" | The vendor's summary of Kar 2024, a one-text study the index records; no test set of its own | undetectable.ai | 9 November 2026 |
| AI or Not | 98.9% (accuracy) via API | No set, date or modality split | aiornot.com | 9 November 2026 |
| Smodin | 99.8% (accuracy) on AI text, 99% on human text | An "Accuracy in Action" section describing three texts; no set or date | smodin.io | 9 November 2026 |
| TruthScan | 99%+ detection (accuracy), qualified in the FAQ as for custom tuned enterprise models; image false positive rate under 1% | No set, date or model for text; no base for the image rate | truthscan.com | 9 November 2026 |
| Codeleaks | 99.8% (accuracy) in the launch release of 9 November 2023; "over 99%" with a 0.2% false positive rate on the live page | A footnote to "internal testing of the English language datasets", the text detector's figure; no code set | copyleaks.com | 9 November 2026 |
| Illuminarty | Validation accuracy 97.5% for binary AI detection, 18 March 2023 | "Our collected dataset", size not stated; the vendor adds the numbers "need not match the actual experience of the users" | illuminarty.ai | 9 November 2026 |
| Hive | "Hive is the clear winner" in "a 2024 independent research study" | A reference to Ha and Passananti 2024, an indexed record; no vendor test set | thehive.ai | 8 November 2026 |
| Sightengine | "the highest accuracy among the tools tested" in a benchmark of 80,000 images | A reference to arXiv 2404.14581, a paper not yet indexed; no vendor test set | sightengine.com | 9 November 2026 |
Figures as quoted on each card from the page read on the date shown.
The remaining 15 cards print no accuracy percentage for their own tool on the pages read: Scribbr, whose page reports its test of other tools (84% for the best premium tool, 68% for the best free one) and no figure for its own (scribbr.com, read 8 November 2026); Walter Writes, GLTR, AIDetectPlus, WasItAI, Reality Defender, Writer, Grammarly Authorship, Turnitin Clarity, SynthID, BrandWell, Surfer, Quetext, Deepware and Sensity. Two of those publish a method paper with a different metric: GLTR's paper reports 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), and SynthID's Nature paper reports detectability as a true positive rate at a fixed 1% false positive rate on the authors' own watermarked text (synthid.com, read 9 November 2026).
A test set a reader cannot obtain is a figure a reader cannot check, whatever the figure says.
Questions about published AI detector test data
Which AI detectors publish their benchmark data?
Twelve of the 72 publish a dated evaluation with a named or sized test set, per the cards read 8 and 9 November 2026; only GPTZero releases predictions for one set, and the one dataset on the record that can be downloaded in full is RAID, a researchers' benchmark with data and code on GitHub (arXiv 2405.07940, read 8 November 2026).
Does GPTZero publish its test data?
Partly: raw predictions for its academic paper review set are published, and the other three English sets of 1,000 human and 1,000 AI texts are available to researchers on request because three of the four "cannot be published" for copyright reasons (gptzero.me, read 9 November 2026).
Does Pangram publish the data behind its 1 in 10,000 claim?
The corpus but not the subset: the 4.1 model card names 781,981 pre-ChatGPT English FineWeb texts, a public corpus, and 520,000 AI generations from 26 models, and does not publish the scored files (pangram.com, read 9 November 2026).
Can anyone download the RAID benchmark?
Yes: Dugan et al. 2024 released the data and code on GitHub and run a leaderboard at raid-bench.xyz, keeping the hidden test labels private so that leaderboard entries cannot be fitted to them (arXiv 2405.07940, read 8 November 2026; RAID record).
Which vendor claims cite a public benchmark they did not run?
Four: QuillBot and Grammarly cite the RAID leaderboard for their 99% figures, Hive cites a 2024 study the index holds as Ha and Passananti 2024, and Sightengine cites arXiv 2404.14581, per their pages read 8 and 9 November 2026. None of the four publishes a test set of its own.
Sources
All read 9 November 2026 unless an earlier date is shown; the vendor claim source of every card is linked in the tables.
- GPTZero AI detection benchmarking, 5 February 2026, updated 20 July 2026: https://gptzero.me/news/gptzero-ai-detection-benchmarking-the-industry-standard-in-accuracy-transparency-and-fairness/
- Pangram 4.1 model card, 19 October 2026: https://www.pangram.com/research/model-card/pangram-4-1
- Pangram 4 model card, 29 July 2026: https://www.pangram.com/research/model-card/pangram-4
- Copyleaks testing methodology: https://copyleaks.com/ai-detector/testing-methodology (read 8 November 2026)
- Originality.ai 5.0 launch post: https://originality.ai/blog/introducing-originality-5-0-and-max (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)
- Compilatio support article on detector reliability: https://support.compilatio.net/hc/en-us/articles/17408769118993
- Proofademic AI text detector: https://proofademic.ai/ai-text-detector/ (read 8 November 2026)
- Winston AI: https://gowinston.ai/ (read 8 November 2026)
- Resemble AI benchmarks: https://www.resemble.ai/benchmarks
- Pindrop Pulse for Contact Centers: https://www.pindrop.com/pulse-for-contact-centers
- GPT-2 Output 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/
- DetectGPT, arXiv 2301.11305: https://arxiv.org/abs/2301.11305
- Dugan et al. 2024, RAID: https://arxiv.org/abs/2405.07940 (read 8 November 2026)
- QuillBot AI detector: https://quillbot.com/ai-content-detector (read 8 November 2026)
- Grammarly AI detector: https://www.grammarly.com/ai-detector (read 8 November 2026)
- Phrasly AI detector: https://phrasly.ai/ai-detector (read 8 November 2026)
- Sapling AI detector: https://sapling.ai/ai-content-detector (read 8 November 2026)
- JustDone AI detector: https://justdone.com/ai-detector (read 8 November 2026)
- ZeroGPT DeepAnalyse: https://www.zerogpt.com/deep-analyse (read 8 November 2026)
- Codequiry AI code detection: https://codequiry.com/detect-ai-written-code
- Undetectable AI detector: https://undetectable.ai/ai-detector
- AI or Not: https://www.aiornot.com/
- Smodin AI detector: https://smodin.io/ai-content-detector
- TruthScan AI detector: https://truthscan.com/ai-detector
- Copyleaks Codeleaks launch release, 9 November 2023: https://copyleaks.com/about-us/press-releases/copyleaks-launches-codeleaks
- Illuminarty release post, 18 March 2023: https://illuminarty.ai/en/posts/big-spring-update.html
- Hive AI-generated content classification: https://thehive.ai/apis/ai-generated-content-classification (read 8 November 2026)
- Sightengine AI image detector: https://sightengine.com/detect-ai-generated-images
- Scribbr AI detector: https://www.scribbr.com/ai-detector/ (read 8 November 2026)
- GLTR paper, arXiv 1906.04043: https://arxiv.org/abs/1906.04043
- SynthID: https://synthid.com/
Revision log: 2026-11-09: page created from the vendor claim and vendor benchmark fields of 42 cards at index version 2026.11.7.; 2026-11-09: eighth-wave pages linked
Related pages: Which AI detectors run on-premise or offline? The four that run on your own machine, the one open-source text component, the one on-device PC tool and the ten vendors that state an on-premise option, Have AI detector scores been used as evidence? What the primary documents in the index state, and what 24 vendors say about acting on a score, AI detector accuracy claims vs published results, ranked by the gap and Questions about AI detectors, answered from the index.