GPTKit AI text detector

Two records in the studies index print results for GPTKit, and they pull in opposite directions. Orenstrakh 2023 ran eight detectors on 114 human and 20 ChatGPT English computing submissions and, by its threshold method, scored GPTKit 100.00% on the human set and 75.00% on the ChatGPT set, 87.50% combined, with 0 false positives in 114; the paper's conclusion names GPTKit "the best detector for avoiding false positives" (arXiv 2307.07411, Tables 6 and 8 and Section 4.1, read 9 November 2026). The same paper's resilience test moved GPTKit from 90% to 30% accuracy on 10 QuillBot-paraphrased ChatGPT texts (Table 9), and on 10 human and 10 ChatGPT submissions in Spanish it classified 80% and 20% correctly (Table 12). The paper also reports that some human submissions scored "up to 37% fake" and that the tool merged separate paragraphs into one.

Perkins 2024 bought credits and ran GPTKit with six other detectors. Its mean accuracy on unmanipulated AI text was 6%, falling to 4.5% after adversarial techniques, a drop of 1.5 points and the lowest post-manipulation figure of the seven, while it made 0 false accusations in 10 human samples (DOI 10.1186/s41239-024-00487-w, Tables 8 and 10, read 9 November 2026). Read together, the two records describe a detector that rarely accuses a human writer and misses most of the AI text it is shown.

The fields follow the index rules. The false positive value is 0 of 114 from Orenstrakh 2023, the larger base, with Perkins 2024's 0 of 10 in the base text. Robustness is 30%, the Orenstrakh 2023 post-QuillBot accuracy, under the rule that takes the highest printed post-attack detection figure; Perkins 2024's 4.5% is quoted beside it. No record prints a miss count, so the false negative ranking carries no GPTKit value, because 75.00% and 6% are percent-correct figures whose complements the rule does not take. The claim gap pairs the vendor's 93% with Orenstrakh 2023's 87.50% combined accuracy, a gap of 5.5 points, with Perkins 2024's 4.5% as the lowest figure in the block.

GPTKit ranks on two study records in independently tested detectors, at 0 in the false positive ranking, at 30% in adversarial robustness, at plus 5.5 points in the claim versus evidence gap, at zero days in data retention and at one record in languages evidence for the Spanish set. On vendor-stated values read 9 November 2026 it qualifies for maximum input per check at 153,600 characters on the Institution plan, languages at one, generator models named at one (ChatGPT), accepted file types, platform integrations on its API, detectors with an API and the no-account view. It also appears on the paraphrased text question and the Spanish question on the Orenstrakh 2023 figures.

Every detector card on the register shows the same fourteen fields, so a detector with 0 false positives in 114 and 4.5% accuracy after manipulation can be read in the same columns as the detectors whose two figures sit the other way round.