Vendor compilation: what Originality.ai's accuracy page and 17-study meta-analysis publish about its own detector

The record joins the studies index beside the GPTZero and Pangram vendor benchmarks and the Turnitin claims record as a vendor-statement record: it documents what one company publishes about itself, dated and versioned, so the claim vs evidence gap ranking can set each claim against the studies that tested the same tool. Vendor figures are never averaged with study figures in this index.

One product, scored by its vendor and, in the vendor's restatement, by 17 outside papers. The meta-analysis post names the detectors each study compared, and the index matches those with cards here (meta-analysis post, Rankings table, read 10 November 2026).

Names as the vendor prints them in the "Key Competitors" column; the vendor restates no competitor's figure in that table (meta-analysis post, read 10 November 2026).

The accuracy page's own figures come from the vendor's model release history, which gives one or two numbers per release with no test-set size: 1.4 (April 2023) a false positive rate of 1.5%; 3.0 Turbo (February 2024) accuracy "from 90.2% to 98.8%" on "our toughest testing dataset"; 1.0.0 Lite (July 2024) 98% (accuracy) "under 1% false positive rate"; Multilingual 2.0.0 (May 2025) 97.8% (accuracy), false negatives 1.99%, false positives 2.4%; 1.0.2 Lite and 3.0.2 Turbo (September 2025) false positive rates of 0.5% and 1.5%; and 5.0 and 5.0 Max (October 2026) 99%+ (accuracy), with the classic models retired in September and October 2026 (Originality.ai accuracy page, Model Release History, read 10 November 2026).

The vendor sizes three of the sets it restates, RAID at 6,287,820 texts across 12 detectors, 11 generators, 8 domains and 11 attacks at a 5% false positive threshold, Walters 2023 at 126 papers and 16 detectors, and the "Empirical Study" at 11,580 samples and 6 tools, and sizes none of its own release tests (RAID post; meta-analysis post, Studies 1 and 2, read 10 November 2026). The index has read RAID and Walters 2023 in the primary source and the other 15 only here.

Every figure is a vendor figure, printed on the page named in its row (Originality.ai accuracy page; meta-analysis post; RAID post, read 10 November 2026).

All values as printed by the vendor on 10 November 2026; the accuracy page repeats the same table.

Two of the 17 studies are in this index. For RAID, the vendor's 85% base accuracy, 96.7% under paraphrase and the two weak attacks match what the RAID record prints from the paper: 85.0 with no attack, 96.7 under paraphrase and 9.3 under homoglyphs at a 5% false positive threshold, with a 0.25% false positive rate at threshold 0.5 that the vendor's post does not quote (RAID post, Findings 1 to 4, read 10 November 2026). For Walters 2023, the vendor restates 97% overall where the Walters record prints 98% correct, 0% incorrect and 2% uncertain on 126 documents from the paper's own table; the per-class figures of 100%, 100% and 95% agree (meta-analysis post, Study 2, read 10 November 2026).

What the compilation leaves out is as dated as what it includes. Five records in this index scored Originality.ai and are absent from the vendor's list: Orenstrakh 2023 with 7 false positives in 114 human coursework texts, Pegoraro 2023 with a 7.6% true positive rate on ChatGPT answers, Liang 2023 with an average 61.22% false positive rate across seven detectors on TOEFL essays, Jabarian and Imas 2025 with a false positive rate of 0.0011 at threshold 0.5, and Epoch AI 2026 with 19 of 495 pre-2022 human passages flagged by Turbo 3.0.2 (meta-analysis post, Rankings table, read 10 November 2026). The vendor's own RAID post also records that "a sub 1% false positive solution is needed", which the Lite model later claimed at 0.5% (RAID post, read 10 November 2026).

The vendor states that "no AI detection tool is ever going to be 100% perfect", that "a single number related to a detector's effectiveness without additional context is useless", and that "many of the tests in this meta-analysis were conducted on legacy models", pointing readers to the 5.0 launch post (Originality.ai accuracy page; meta-analysis post, read 10 November 2026). The RAID post notes the study "did not include the domain/type of content we at Originality.ai focus most on, i.e. web content".

Evident from the pages: the vendor selected which studies to list and which figure to headline from each; the release-history figures name no test-set size; the meta-analysis "Key Findings" line counts "all 6 published 3rd party studies" while the table lists 17; the Walters figure differs from the paper by one point; and the accuracy page is the source of the claim the Originality.ai card quotes.

One card takes a dated vendor-compilation line, and no competitor card takes a figure, because the vendor restates none (Originality.ai accuracy page; meta-analysis post, read 10 November 2026).

None of these figures enters the false positive rate ranking or the false negative rate ranking, which take study results only; the claim vs evidence gap ranking is where the "99%+" claim meets Walters 2023, RAID and Epoch AI 2026. How this record is sourced explains the rule, and the home page shows which AI detectors exist and who has tested them.

Related records: the GPTZero and Pangram vendor benchmarks record scores Originality lite-102 at a 1.88% false positive rate on a competitor's set; Epoch AI 2026 and RAID 2024 are the outside tests the vendor's claim can be set against.