AI detector accuracy claims vs published results, ranked by the gap

Two pages added 9 November 2026 read the claims this gap starts from: the published test data question, which separates the twelve vendors whose evaluation names or sizes its test set from the fifteen with a figure and no set, and the generators covered ranking, which counts the models each vendor names, another claim the studies do not test.

Field sorted: the vendor's headline accuracy percentage minus the highest accuracy an indexed study prints for the tool on AI-generated or mixed text, in percentage points, compiled on 9 November 2026. The claim is the first accuracy percentage on the vendor's product page as quoted on the card; "over 99%" and "99%+" are read as 99. The study figure is the highest the paper prints as an accuracy or percent correct on an AI or mixed sample, with its base and setting; a specificity on human text alone, an AUROC or a true positive rate is not used; where a card's only record scored images and the claim covers images, as for AI or Not, the image accuracy is paired and the row marks it. Direction: lower ranks higher, so a claim below its best result ranks above a claim above it; a negative gap is not a judgement that the claim is modest, only that one study printed a higher figure. Tie rule: Copyleaks and Originality.ai tie at minus 1, so ties break on study records per card, most first (6 to 5), then product name alphabetically. Qualification rule: a numeric accuracy claim about the vendor's own tool and at least one indexed study that prints an accuracy figure for it; the second table names the cards that fail one side or the other. 7 of 22 qualified on the morning of 9 November 2026; 8 of 28 after the third wave of cards, AI or Not joining with the one image pairing; 9 of 72 after the fourth wave, Compilatio joining at plus 15 points. Where a vendor states a range, as Compilatio's 94 to 99% does, the lower figure is paired, so the gap is the smallest the range allows. Full rules: methodology page; every record: studies index. The same pairing for false positive claims, nine vendors stating under 1% beside the studies that checked them, is on the question page on sub-1% claims, and the question page on ChatGPT coverage marks which generators each claim names.

A vendor's accuracy and a study's accuracy are rarely the same quantity. GPTZero's 99.9% and Grammarly's and QuillBot's 99% all cite RAID, yet the RAID paper prints accuracy at a fixed 5% false positive rate per generator and decoding setting, from 98.8% to 9.4% for GPTZero (arXiv 2405.07940, Table 5, read 8 November 2026), and does not test Grammarly or QuillBot. Walters 2023 counted correct verdicts on 126 essays with an "uncertain" option (DOI 10.1515/opis-2022-0158); Van Vlasselaer 2026 scored 4,000-word papers with strict and inclusive bands (DOI 10.1007/s40979-026-00226-w, read 9 November 2026); Copyleaks' own methodology page measures human-only accuracy on 326,120 texts (copyleaks.com, read 8 November 2026). The version differs too: Walters tested Originality.ai model 1.4 and the claim is for 5.0. The gap on this page is a distance between two published numbers, each shown with its base so a reader can see what was measured.

Since 9 November 2026 the count of cards quoting an accuracy figure in the lead counts a percentage figure in the claim field, a digit followed by a percent sign, rather than any digit, so that dates and ranges no longer count a card; the model card question, added the same day, applies the stricter test of a sized test set and finds 8 cards.