Is there a legit AI detector? How to tell from the record

The first check is a method: what was counted, on how many texts, at which threshold, on which date. Copyleaks publishes a methodology page with a test date of 10 September 2026, a human-only sample of 326,120 texts and a default false positive rate of 0.30% (copyleaks.com, read 8 November 2026); the same vendor's product page states .03%, and the card quotes both. Turnitin publishes the base for its false positive claim, over 700,000 pre-ChatGPT papers per release, and no accuracy or recall figure (Turnitin FAQ, read 8 November 2026). A vendor figure with a base can be checked; one without cannot.

The second check is a study record from outside the company. Ten detectors in the index have one or more: GPTZero 7, ZeroGPT 5, Turnitin 4, Copyleaks 4, the GPT-2 Output Detector 4, Originality.ai 3, Winston AI 2, Sapling 2, GLTR 2 and Scribbr 1, as counted in the ranking by independent testing on 9 November 2026. The other twelve cards, QuillBot, Grammarly, Pangram, Phrasly, JustDone, Proofademic, Walter Writes, Hive, AIDetectPlus, Codequiry, WasItAI and Reality Defender, have no study record, which says only that no indexed paper has put those tools on a shared sample.

The third check is a false positive rate with a base, because a tool that flags human writing does more harm than a tool that misses AI writing. Eight of the ten tools with a record posted 0 false positives in at least one study, per the false positive rate ranking, and the same tools posted other figures elsewhere: Copyleaks 0 of 42 in Walters 2023 (DOI 10.1515/opis-2022-0158) and 5 of 10 in Perkins 2024 (DOI 10.1186/s41239-024-00487-w). Two vendors without a study record publish a rate with a base: Pangram states 1 in 10,000 on public datasets of tens of millions of documents (pangram.com, read 8 November 2026), and Proofademic states 0.2% on 4,200 human-written academic texts (proofademic.ai, read 8 November 2026); both sit in the vendor table of that ranking, unranked.

The fourth check is what happens to the text. Eighteen of the 22 cards carry a retention statement from the vendor's policy or docs, read 8 and 9 November 2026, and they differ: Scribbr says submissions are not stored or shared; GPTZero stores dashboard inputs and not API inputs; Copyleaks and Originality.ai may use submissions to train models unless the customer opts out. The methodology page lists the policy page each statement was read from.

Vendor language runs in two directions. Nine cards quote an accuracy figure of 98% or above, and every card carries a caveat in the vendor's own voice, six of them quoted here: QuillBot says "Scores are signals, not verdicts"; Scribbr says "we can't guarantee 100% accuracy"; Phrasly says "no method of detection is 100% accurate"; JustDone says "Every AI detection result is a probability, not a verdict"; Walter Writes says "No AI detector can guarantee 100% accuracy"; Turnitin says the indicator "should not be used as the sole basis for action", each quoted on its card from the vendor page read 8 November 2026. A caveat speaks to the vendor's candour, not to the tool's accuracy.

A label is not a study. JustDone marks its 80% and 10.3% figures "Independently benchmarked" while the vendor's own blog describes the vendor's own evaluation framework (justdone.com, read 8 November 2026). GPTZero, Grammarly and QuillBot each cite RAID for their accuracy statements; RAID's Table 5 gives GPTZero 98.8% in one setting and 9.4% in another (DOI 10.18653/v1/2024.acl-long.674), and Grammarly and QuillBot are not among the 12 detectors in the RAID paper, listed on the RAID record.

As one measurement on one date. GPTZero recorded 0.00% false positives at threshold 0.5 in RAID 2024 and 9 false accusations in 18 human documents in Weber-Wulff 2023, 6 of them on machine-translated human text (DOI 10.1007/s40979-023-00146-z). Liang 2023 reports an average false positive rate of 61.22% across seven detectors on 91 essays by non-native English writers (DOI 10.1016/j.patter.2023.100779). A tool that is legit on the record can still be wrong on the text in front of you; the record tells you how often. The home board shows which AI detectors exist and who has tested them, and the studies index holds the nine records.

All read 8 November 2026.

• Weber-Wulff et al. 2023: https://doi.org/10.1007/s40979-023-00146-z
• Dugan et al. 2024 (RAID): https://doi.org/10.18653/v1/2024.acl-long.674
• Walters 2023: https://doi.org/10.1515/opis-2022-0158
• Perkins et al. 2024: https://doi.org/10.1186/s41239-024-00487-w
• Liang et al. 2023: https://doi.org/10.1016/j.patter.2023.100779
• Chaka 2023: https://doi.org/10.37074/jalt.2023.6.2.12 (read 9 November 2026)
• Copyleaks testing methodology: https://copyleaks.com/ai-detector/testing-methodology
• Copyleaks privacy policy: https://copyleaks.com/privacy-policy
• Turnitin AI writing detection FAQs: https://guides.turnitin.com/hc/en-us/articles/28477544839821
• GPTZero privacy policy: https://gptzero.me/privacy-policy.html
• GPTZero blog, 25 October 2026: https://gptzero.me/news/best-ai-detectors
• ZeroGPT privacy policy: https://www.zerogpt.com/privacy-policy
• ZeroGPT pricing: https://www.zerogpt.com/pricing
• Originality.ai privacy policy: https://originality.ai/privacy-policy
• Winston AI privacy policy: https://gowinston.ai/privacy-policy/
• Sapling privacy policy: https://sapling.ai/privacy
• Sapling pricing: https://sapling.ai/pricing
• Scribbr AI detector: https://www.scribbr.com/ai-detector/
• Pangram: https://www.pangram.com/
• Proofademic AI text detector: https://proofademic.ai/ai-text-detector/
• JustDone AI detector: https://justdone.com/ai-detector
• GLTR repository: https://github.com/HendrikStrobelt/detecting-fake-text (read 9 November 2026)
• GPT-2 Output Detector repository: https://github.com/openai/gpt-2-output-dataset/tree/master/detector (read 9 November 2026)

Revision log: 2026-11: page created; 2026-11-09: 22-card recount; 2026-11-09: wave 2 records noted, 12 of 22 with a record