Which AI detectors were in the Weber-Wulff 2023 study? The 14 tools, their result on 54 documents, and what each one is today
Fourteen AI text detectors were tested in Weber-Wulff et al. 2023, "Testing of detection tools for AI-generated text", on 54 documents between 7 March and 8 May 2023, and 9 of the 14 have a card in this index of 72 detectors, by the roster on the Weber-Wulff 2023 record read 9 November 2026 (DOI 10.1007/s40979-023-00146-z, read 8 November 2026). The 14 are Check For AI, Compilatio, Content at Scale, Crossplag, DetectGPT, Go Winston, GPT Zero, the GPT-2 Output Detector Demo, the OpenAI Text Classifier, PlagiarismCheck, Turnitin, the Writeful GPT Detector, Writer and ZeroGPT. Turnitin scored highest at 41 of 54 documents correct (76%) and Content at Scale lowest at 18 of 54 (33%), and three and a half years on, two of the 14 have been withdrawn by their makers, one has been renamed, one hosted demo is paused and one has vanished without a statement (shut-down detectors question, read 9 November 2026).
The study is the one most often cited when a reader meets the claim that "AI detectors are under 80% accurate", so the table lists every tool it tested with the three figures the paper prints per tool, binary accuracy over 54 documents, false accusations over 18 human documents and undetected cases over 36 AI documents (Tables 5, 9 and 10), beside what the index knows about the tool at the read date. A status is recorded only where a vendor page or the shut-down question states it; a tool the index has not carded is marked "not checked", because the index records what was read and not what it infers. This page is one of the questions answered from the index; the ranking by independent testing counts this record for 9 cards, and the AI detector register carries every card.
| Tool, as the paper names it | Index card | Binary accuracy (54 documents) | False accusations (18 human) | Undetected (36 AI) | Status at 9 November 2026 |
|---|---|---|---|---|---|
| Turnitin | Turnitin | 41, 76%, rank 1 | 0, 0.0% | 10, 27.8% | Live; card read 8 November 2026 |
| Compilatio | Compilatio | 40, 74%, rank 2 | 0, 0.0% | 11, 30.6% | Live as Magister+; card read 9 November 2026 |
| GPT-2 Output Detector Demo | GPT-2 Output Detector | 39, 72%, rank 3 | 2, 11.1% | 12, 33.3% | Hosted demo paused; code and weights public under the MIT License |
| Crossplag | no card | 37, 69%, rank 4 | 3, 16.7% | 13, 36.1% | Not checked |
| Go Winston | Winston AI | 36, 67%, rank 5 | 0, 0.0% | 12, 33.3% | Live as Winston AI; card read 8 November 2026 |
| Check For AI | Check For AI | 32, 59%, rank 6 | 1, 5.6% | 12, 33.3% | Withdrawn: sunset notice from the Wayback capture of 23 July 2023 |
| ZeroGPT | ZeroGPT | 32, 59%, rank 6 | 0, 0.0% | 13, 36.1% | Live; card read 8 November 2026 |
| GPT Zero | GPTZero | 29, 54%, rank 8 | 9, 50.0% | 3, 8.3% | Live as GPTZero; card read 8 November 2026 |
| OpenAI Text Classifier | no card (record) | 29, 54%, rank 8 | 0, 0.0% | 16, 44.4% | Withdrawn 20 July 2023 "due to its low rate of accuracy" |
| Writer | Writer | 27, 50%, rank 10 | 1, 5.6% | 19, 52.8% | Withdrawn 22 December 2025 with its API endpoint |
| DetectGPT | DetectGPT | 25, 46%, rank 11 | 0, 0.0% | 13, 36.1% | Hosted demo gone by 26 November 2023, no statement; MIT-licensed code on GitHub |
| Writeful GPT Detector | Writefull | 23, 43%, rank 12 | 1, 5.6% | 18, 50.0% | Discontinued: notice on the live route since the bundle captured 7 March 2025 |
| PlagiarismCheck | PlagiarismCheck | 21, 39%, rank 13 | 0, 0.0% | 19, 52.8% | Live product, card read 9 November 2026 |
| Content at Scale | BrandWell | 18, 33%, rank 14 | 0, 0.0% | 36, 100.0% | Renamed BrandWell in 2024; detector page redirects to a Moxby Mod since 29 August 2026 |
Figures as printed in Weber-Wulff et al. 2023, Tables 5, 9 and 10 (arXiv 2306.15666, read 9 November 2026; the record page carries 10 of the 14 binary rows and the Crossplag false accusation count, and the Check For AI, Crossplag, PlagiarismCheck and Writeful rows were taken from the paper on 9 November 2026). Statuses from the card pages and the shut-down detectors question, read 9 November 2026.
Which of the 14 Weber-Wulff tools still exist?
Five run today as live products, one runs under a new name, two run only from their published code, two are withdrawn and four are unchecked, by the cards and the shut-down question read 9 November 2026. Turnitin, Compilatio, Winston AI (tested as Go Winston), ZeroGPT and GPTZero are live products with cards read on 8 and 9 November 2026 (Turnitin, Compilatio, Winston AI, ZeroGPT, GPTZero). Content at Scale became BrandWell in 2024 and its detector page has redirected to the AI Content Detector Mod on the Moxby marketplace since 29 August 2026, a move rather than a withdrawal (BrandWell card, read 9 November 2026). The OpenAI Text Classifier was withdrawn on 20 July 2023 "due to its low rate of accuracy", four months after the test window closed (openai.com, read 8 November 2026), and Writer's detector was sunset on 22 December 2025 together with its API (archived writer.com page, captured 9 December 2025, read 9 November 2026). The GPT-2 Output Detector's hosted demo, the address the study used, shows a paused state while the MIT-licensed code and weights stay public (github.com, read 9 November 2026), and the DetectGPT demo at detectgpt.ericmitchell.ai has returned 404 since a Wayback capture of 26 November 2023, with no statement from its authors (github.com, read 9 November 2026). Check For AI, Crossplag, PlagiarismCheck and the Writeful GPT Detector had no card on 9 November 2026, so their status was not recorded. On 10 November 2026 three of the four gained cards: PlagiarismCheck is a live product with six LMS integrations (PlagiarismCheck card, read 9 November 2026); Check For AI was withdrawn in 2023, with the banner "We will be sunsetting CheckForAi soon" in the Wayback capture of 3 June 2023 and "We have sunset the service" from 23 July 2023 onward (Check For AI card, read 10 November 2026); and the Writeful GPT Detector, which the vendor spells Writefull, was discontinued between the captures of 31 December 2024 and 7 March 2025, its live route now reading "we've decided to take our GPT Detector offline" (Writefull card, read 10 November 2026). That leaves four of the 14 withdrawn by their makers (the OpenAI classifier, Writer, Check For AI and Writefull) and only Crossplag and the OpenAI classifier without a card.
What did each tool score in Weber-Wulff 2023?
Binary accuracy ran from 41 of 54 for Turnitin to 18 of 54 for Content at Scale, and the two counts that matter most to a reader pulled in opposite directions (arXiv 2306.15666, Tables 5, 9 and 10, read 9 November 2026). GPT Zero made 9 false accusations in 18 human documents, 50.0%, the highest of the 14, and missed only 3 of 36 AI documents, 8.3%, the lowest. Content at Scale made no false accusations and missed all 36 AI documents. Eight tools made no false accusation at all: Turnitin, Compilatio, Go Winston, ZeroGPT, the OpenAI Text Classifier, DetectGPT, PlagiarismCheck and Content at Scale. Six made at least one, and the machine-translated class 02-MT drove most of them: Crossplag's 3 accusations, the GPT-2 Output Detector Demo's 2 and the single accusations of Check For AI, Writeful and Writer all fell in that class, and 6 of GPT Zero's 9 did. On the AI side, Turnitin was the only tool to classify all 18 unedited ChatGPT documents correctly, and its 10 misses were all among the manually edited and QuillBot-paraphrased texts (Weber-Wulff 2023 record). Averaged across the 14 tools, 71.4% of the paraphrased documents went undetected, the figure the paraphrased text question carries for the whole roster.
The record page prints 10 of the 14 binary rows and the Crossplag false accusation count; the Check For AI, Crossplag, PlagiarismCheck and Writeful GPT Detector cells above were read from Tables 5, 9 and 10 of the preprint on 9 November 2026 so that every tool carries the same three fields. Nothing on this page comes from a vendor's own account of the study.
Does a 2023 Weber-Wulff result describe the tool today?
No, for two reasons the paper itself gives and one the dates add. The authors restricted their claims to English, left computer code out, noted that results from the same tool on the same text varied over time, and disclosed that Turnitin and PlagiarismCheck knew the documents were a test suite (DOI 10.1007/s40979-023-00146-z, read 8 November 2026). Since the March to May 2023 window, Turnitin's release notes list model changes on 24 May 2023, 6 December 2023, 16 July 2024, 27 August 2025, 14 October 2025, 12 February 2026 and 14 October 2026 (guides.turnitin.com, read 9 November 2026), GPTZero's leaderboard entry on RAID is dated 7 October 2025 and Winston AI's 24 September 2026 (leaderboard question), and four of the 14 tools cannot be run the way the study ran them: two are withdrawn and two survive only as code. The repeatability question sets the authors' replicability note beside the other records that address repeated runs.
Where the nine carded tools were tested again
| Index card | Records after Weber-Wulff 2023 | One later record and its human-text result | Source |
|---|---|---|---|
| GPTZero | 15 more, 16 in all | Epoch AI 2026: 0 of 495 false positives for model 2026-05-11-base | Epoch AI 2026 record |
| GPT-2 Output Detector | 11 more, 12 in all | Jabarian and Imas 2025: 0.9503 false positive rate on 1,992 human passages | Jabarian and Imas 2025 record |
| ZeroGPT | 6 more, 7 in all | RAID 2024: 1.42% false positive rate at threshold 0.5 on human documents | RAID record |
| Turnitin | 4 more, 5 in all | Van Vlasselaer 2026: all 40 human papers classified correctly | Van Vlasselaer 2026 record |
| Writer | 4 more, 5 in all | Walters 2023: 14% of 42 human papers called AI, the joint highest of 16 tools | Walters 2023 record |
| DetectGPT | 3 more, 4 in all | Sadasivan 2023: false positive rate fixed by threshold | Sadasivan 2023 record |
| BrandWell | 2 more, 3 in all | Walters 2023: 10% of 42 human papers called AI, as Content at Scale | Walters 2023 record |
| Winston AI | 1 more, 2 in all | RAID 2024: 0.55% false positive rate at threshold 0.5 | RAID record |
| Compilatio | 0 more, 1 in all | Weber-Wulff 2023 is the only record | Weber-Wulff 2023 record |
Record counts from each card's study list in the data file, read 9 November 2026; results as each record prints them. The five tools without a card have no later record in the index.
A result from spring 2023 describes 54 documents and 14 tools as they stood that spring, and two of those tools cannot be run on a 2026 document at all, while two more run only from their published code.
Questions about the Weber-Wulff 2023 roster
How many tools did Weber-Wulff 2023 test?
14, chosen from 18 considered, on 54 documents in six classes of 9, for 756 tests between 7 March and 8 May 2023 (DOI 10.1007/s40979-023-00146-z, read 8 November 2026). 9 of the 14 had a card in this index on 9 November 2026 and 12 have one on 10 November 2026, after the PlagiarismCheck, Check For AI and Writefull cards joined; only Crossplag and the OpenAI classifier, which has its own record, remain without one.
Which Weber-Wulff 2023 tools no longer exist?
2 were withdrawn by their makers, the OpenAI Text Classifier on 20 July 2023 (openai.com, read 8 November 2026) and Writer on 22 December 2025 (archived writer.com page, read 9 November 2026); the GPT-2 Output Detector demo is paused and the DetectGPT demo gone by 26 November 2023, with the code of both still public. Two more were withdrawn without a dated vendor statement beyond the notice itself: Check For AI, sunset by the Wayback capture of 23 July 2023, and Writefull's GPT Detector, discontinued by 7 March 2025 (Check For AI card, Writefull card, read 10 November 2026).
Did Weber-Wulff 2023 test GPTZero or Turnitin?
29 of 54 documents (54%) for GPT Zero, with 9 false accusations in 18 human documents, and 41 of 54 (76%) for Turnitin, with 0 false accusations, in Weber-Wulff 2023 (arXiv 2306.15666, Tables 5 and 9, read 9 November 2026). Turnitin was tested from 14 to 20 April 2023 and knew the documents were a test suite.
Is Content at Scale the same as BrandWell?
18 of 54 correct and 36 of 36 AI documents undetected is the Weber-Wulff 2023 result of the detector tested as Content at Scale, which is sold as BrandWell since 2024, and its detector page has redirected to a Moxby Mod since 29 August 2026 (BrandWell card, read 9 November 2026).
Where is Crossplag today?
37 of 54 (69%) with 3 false accusations is Crossplag's result in Weber-Wulff 2023 (arXiv 2306.15666, read 9 November 2026); the tool has no card in this index and no vendor page for it has been read, so the index records no status.
Sources
- Weber-Wulff et al. 2023, journal version: https://doi.org/10.1007/s40979-023-00146-z (read 8 November 2026)
- Weber-Wulff et al. 2023, preprint with Tables 5, 9 and 10: https://arxiv.org/pdf/2306.15666 (read 9 November 2026)
- OpenAI, New AI classifier for indicating AI-written text, with the 20 July 2023 update: https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/ (read 8 November 2026)
- Writer AI Content Detector, archived 9 December 2025: http://web.archive.org/web/20251209115654/https://writer.com/ai-content-detector/ (read 9 November 2026)
- GPT-2 output dataset repository and detector README: https://github.com/openai/gpt-2-output-dataset/tree/master/detector (read 9 November 2026)
- DetectGPT repository: https://github.com/eric-mitchell/detect-gpt (read 9 November 2026)
- Moxby AI Content Detector Mod (BrandWell): https://moxby.com/marketplace/ai-content-detector/ (read 9 November 2026)
- Turnitin AI writing detection model release notes: https://guides.turnitin.com/hc/en-us/articles/28294949544717-AI-writing-detection-model (read 9 November 2026)
Revision log: 2026-11-09: page created from the Weber-Wulff 2023 record, the preprint's Tables 5, 9 and 10 and the card pages at index version 2026.11.8; 2026-11-10: Check For AI, Writefull and PlagiarismCheck rows linked to their cards with statuses (tenth wave; the PlagiarismCheck card dates from the ninth); 12 of 14 carded, four withdrawn, Crossplag and the OpenAI classifier the two without a card
Related pages: Which AI detectors detect AI voice and video in one product? The nine cards that list both, and the zero records that tested either, 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, AI detectors ranked by independent testing Two records added in the fifth wave on 9 November 2026, Pegoraro 2023 and Fleckenstein 2024, changed nine counts and none: Pegoraro 2023 tested ZeroGPT, GPTZero, Copyleaks, Originality.ai, Writer, Content at Scale (BrandWell), DetectGPT, the GPT-2 Output Detector and GLTR, lifting GPTZero to 12, the GPT-2 Output Detector to 10, Copyleaks to 7 and Originality.ai and ZeroGPT to 6; Fleckenstein 2024 tested 289 teachers and no software, so it counts for no card ([Fleckenstein 2024 record](/studies/fleckenstein-2024-teachers/)). Illuminarty, one of four cards added in the same wave, joined with its one image record from Ha and Passananti 2024, so 18 of 72 cards now qualify. and Questions about AI detectors, answered from the index.