AI detector false positive rates, ranked

Field sorted: the published false positive rate on each card, from the independent study reporting the lowest rate for that detector, compiled on 9 November 2026. Direction: lower ranks higher. Tie rule: twelve tools tie at 0, so ties break on study records per card, most first, then product name alphabetically (GPTZero 12, the GPT-2 Output Detector 10, Copyleaks 7; Originality.ai and ZeroGPT 6 each; Turnitin 5; DetectGPT 4; BrandWell 3; Pangram, Sapling and Winston AI 2 each; Compilatio 1, counts as of the fifth wave on 9 November 2026, when Pegoraro 2023 lifted nine of them and moved ZeroGPT above Turnitin). Qualification rule: a rate from an independent study in the index; vendor-stated rates sit in a separate table, unranked, and cards with no published rate are omitted. 23 of 72 qualified; the GPT-2 Output Detector and GLTR were added on 9 November 2026, Pangram moved into the ranking later the same day when the Jabarian and Imas 2025 record was indexed, and Writer joined at 5.6% with the third wave of cards the same day; DetectGPT, BrandWell and Compilatio joined at 0 from Weber-Wulff 2023 with the fourth wave. TruthScan publishes a "False positive rate <1%" for image detection with no base, an image figure kept off this text-only field and carried on the TruthScan card. Hive's 0% on 280 human artworks (Ha and Passananti 2024) is an image result and is listed in the false negative rate ranking and the annual false positive evidence review, not here, because this page sorts rates on human-written text; AI or Not's 24.47% on the same 280 artworks, measured under the name Optic in the same paper, is kept off this page by the same rule and sits on the AI or Not card. Full rules: methodology page; each paper: studies index. The published false positive rate view lists every card with a figure of either kind in one table, and the question on vendor claims under 1% sets each vendor's claim beside the studies that checked it.

The false positive studies question, added 9 November 2026, counts how many records stand behind each figure here: GPTZero's lowest rate is one of 11 human-text figures across its records, Scribbr's and Compilatio's one of one.

A rate depends on the texts, the threshold and the date. Weber-Wulff 2023 used 18 documents, 9 machine-translated, where 6 of GPTZero's 9 false accusations fell (arXiv 2306.15666). Walters 2023 allowed an uncertain verdict, which is why Sapling shows 0 false positives with 31% uncertain (DOI 10.1515/opis-2022-0158). RAID 2024 reports four rates per tool at thresholds 0.25 to 0.95 (arXiv 2405.07940). Copyleaks publishes 0.03% on the product page and 0.30% on the methodology page (copyleaks.com). All read 8 November 2026. Jabarian and Imas 2025 fixed thresholds on each API's own probability field, which is why GPTZero's 0.0071 holds at every threshold while Pangram's falls from 0.0010 to 0.0000 (NBER w34223, read 9 November 2026); Orenstrakh 2023 counted a false positive at a 50% probability on the authors' scale (arXiv 2307.07411, read 9 November 2026).