Kar et al. 2024: how sensitive are the free AI detector tools, ten tools on one ChatGPT article and three paraphrases

This record joins the studies index with its smallest text sample: one AI-written article and three machine paraphrases of it, with no human-written text. The journal and DOI are confirmed from the PubMed Central full text (PMC 11572508, read 9 November 2026 through the Europe PMC full-text service).

Kar 2024 took the first ten results of a Google search for "AI detector free" that offered a free tier accepting at least 500 words in one run. Eight of the ten have a card in this index, the Content Detector AI card added on 10 November 2026 with the ninth wave and the Wordtune card later the same day with the tenth (DOI 10.1177/02537176241247934, Methods and Table 1, read 9 November 2026).

Detectors as named in Kar et al. 2024, DOI 10.1177/02537176241247934, Table 1, which lists each tool's URL. Free tiers were used throughout; the paid tiers were not tested.

The text was a brief scientific article on the role of electroconvulsive therapy in treatment-resistant depression, generated by ChatGPT 3.5 on 11 January 2024 from a prompt asking for under 500 words. The same text was then rephrased three ways: QuillBot free in its "standard" setting, Grammarly premium with the "improve it" command, and ChatGPT 3.5 with the prompt "Kindly rephrase the following content". All four texts were scored by the ten tools on 12 January 2024 (DOI 10.1177/02537176241247934, Methods, read 9 November 2026).

Sample as described in Kar et al. 2024, DOI 10.1177/02537176241247934, Methods.

The one metric is the AI-content percentage each tool displayed. The investigators treat the whole text as 100% AI and read a tool's score as its sensitivity on that text; the complementary "human content" figure is also printed for the original article. No threshold, repeat runs or statistics beyond percentages are reported (DOI 10.1177/02537176241247934, Methods and Statistical Analysis, read 9 November 2026).

Every figure below is a cell of Table 1 in Kar 2024; the base in every row is one text (DOI 10.1177/02537176241247934, Table 1, read 9 November 2026).

Values as printed in Kar et al. 2024, DOI 10.1177/02537176241247934, Table 1; the paraphrase rows are printed without percent signs in the table.

The spread on the original article is the main finding: on one entirely machine-written text, five tools said 100% AI, two said 95% or more, and three said 78.26%, 61% and 0%. The authors describe sensitivity ranging "from 0% to 100%" and single out Dupli Checker for "inaccurately detecting content to be 100% human generated" (DOI 10.1177/02537176241247934, Results and Discussion, read 9 November 2026).

The QuillBot paraphrase did the most damage. It took Copyleaks and Wordtune from 100% AI to 0, ZeroGPT from 95.03% to 32.31 and GPTZero from 97% to 29, while Sapling and Undetectable AI stayed at 100 on all three paraphrases (DOI 10.1177/02537176241247934, Table 1, read 9 November 2026). The Grammarly and ChatGPT paraphrases moved most tools less: five tools read both at 100, and GPTZero dropped to 38 on the Grammarly version but held 97 on the ChatGPT version.

The paper's summary sentences do not all match Table 1. The abstract and the Key Messages say Sapling and Undetectable AI read all three paraphrases at 100%, which the table supports; the Discussion says "only three AI-content detectors (QuillBot, Sapling, and Undetectable AI)" held through rephrasing, while the table prints QuillBot at 90 on its own paraphrase (DOI 10.1177/02537176241247934, Key Messages, Discussion and Table 1, read 9 November 2026). This record follows the table.

The authors name two limits: the comparison covers only ten free versions, and it uses "a limited number of AI-generated and paraphrased contents"; they add that subscription tiers "may be more sensitive" and that sensitivity "may change over time with software upgrades" (DOI 10.1177/02537176241247934, Discussion and Conclusion, read 9 November 2026). The paper is a four-page report with no human text, one topic, one generator and one run per tool, so nothing in it measures false positives and the scores cannot be averaged into an accuracy. The repeatability question, added 9 November 2026, reads that sentence beside the other records that note change over time.

Eight detectors in Kar 2024 have cards in this index; the Content Detector AI card joined on 10 November 2026 with the ninth wave of cards and records its 78.26% AI reading on the original article and 64.58 after the QuillBot paraphrase, and the Wordtune card joined later the same day with the tenth wave and records 100% AI on the original article, 0 after the QuillBot paraphrase and 100 after the Grammarly and ChatGPT paraphrases, the only record of a tool whose page has returned 404 since July 2025. Each card records the tool's Table 1 score on the original ChatGPT article and its score after the QuillBot paraphrase, the two cells that separate the tools (DOI 10.1177/02537176241247934, Table 1, read 9 November 2026).

All eight appear in the ranking of detectors by test record. No figure from this record enters the false positive rate ranking, which takes only rates printed on human text, or the false negative rate ranking, which takes printed rates on sets of AI texts and never a complement of a sensitivity score; a single-text reading is neither. The paraphrase cells are noted on each card for the adversarial robustness ranking as single-text observations, not rates. The home page lists which AI detectors exist and who has tested them.

Related records: Walters 2023 ran 16 detectors, Copyleaks, ZeroGPT, GPTZero and Sapling among them, on 126 documents and is the comparison the authors draw; Elkhatat 2023 is their other comparison, five tools on 15 ChatGPT paragraphs and five human controls; Perkins 2024 measures paraphrase and other adversarial edits on a larger sample (DOI 10.1177/02537176241247934, Discussion, read 9 November 2026).

Sapling, which held 100 on every paraphrase here, and GPTZero are compared on this and two other shared records in the Sapling vs GPTZero comparison, read 10 November 2026.