AI detector false negative rates, ranked

Field sorted: the lowest share of AI-generated text an indexed study prints as missed, undetected or labelled human for each detector, compiled on 9 November 2026. Direction: lower ranks higher. Tie rule: Copyleaks, Originality.ai and Turnitin tie at 0, the GPT-2 Output Detector and Winston AI at 33.3%, and ZeroGPT and DetectGPT at 36.1%, so ties break on study records per card, most first, then product name alphabetically. Qualification rule: the study prints the rate or a count of missed AI texts with its base, at the threshold the study used; a complement of a sensitivity or "percent correct" figure is not taken, which is why Scribbr (69% correct on 42 GPT-4 papers) and Sapling (33% correct) do not qualify and Gao 2023's 86% sensitivity feeds no value. Where a study prints the rate per generator and no pooled figure, as Jabarian and Imas 2025 do, the lowest per-generator rate at the 0.5 threshold is taken and the others are quoted in the block. The false positive rate from the same study stands in every row, because a low miss rate means nothing beside a high flag rate. 10 of 22 qualified on the morning of 9 November 2026; 12 of 28 qualified after the third wave of cards, AI or Not at 1.15% and Writer at 40%; 15 of 38 qualified after the fourth wave, BrandWell at 15%, Compilatio at 30.6% and DetectGPT at 36.1%; 16 of 72 qualify after the fifth wave, Illuminarty at 4.69% on images from Ha and Passananti 2024. Pegoraro 2023, added in the same wave, prints true positive rates for nine cards, and 100 minus a true positive rate is a complement this rule does not take, so no Pegoraro figure enters the field. Full rules: methodology page; every record: studies index.

The GPT-4 question, added 9 November 2026, reads the GPT-4.1 cells of Jabarian and Imas 2025 that this ranking prints for Pangram, GPTZero and Originality.ai, which the record now carries as a row of its own.

Walters 2023 fed 84 essays generated with GPT-3.5 and GPT-4 to each tool in June and July 2023 and counted a miss as an "incorrect" verdict (DOI 10.1515/opis-2022-0158); Weber-Wulff 2023 counted documents in six classes, three of them edited, translated or paraphrased, so its 36 AI documents were harder than Walters's (arXiv 2306.15666); Perkins 2024 built 104 samples with adversarial prompting (arXiv 2403.19148), read 8 November 2026. Van Vlasselaer 2026 generated 4,000-word papers with ChatGPT Deep Research mode and counted a fully AI paper under 20% as a false negative, which produced 100% for Turnitin and 70% for GPTZero (DOI 10.1007/s40979-026-00226-w); Jabarian and Imas 2025 fixed thresholds on each API's probability field over 1,992 passages per model (NBER w34223), read 9 November 2026. Copyleaks and Turnitin can post 0 of 84 in one paper and 75% and 100% of 40 in another because the text, the generator, the length and the year all changed.