Walters 2023: the effectiveness of 16 AI text detectors on student essays

This record is one of eight in the studies index. The detector tests ran from June 25 through July 12, 2023, and the AI essays were generated in the first week of April 2023 (publisher page).

Walters 2023 lists 16 detectors in Table 3 of the paper, tested June 25 to July 12, 2023. Six required a subscription or payment: Copyleaks, Turnitin, Originality.ai, GPTZero (Pro plan), GPT Radar and Sapling; Originality.ai ran detection model 1.4 (DOI 10.1515/opis-2022-0158).

Detectors as listed in Walters 2023, Table 3, DOI 10.1515/opis-2022-0158.

The sample was three sets of 42 documents. Each of the 126 documents went to each of the 16 detectors, for 2,016 responses in Walters 2023 (publisher PDF).

Sample as described in Walters 2023, DOI 10.1515/opis-2022-0158.

In Walters 2023, front matter, tables, figures and works cited were stripped, and documents were cut to each detector's maximum length. Every response was coded AI, human or uncertain; the false positive rate is the share of the 42 human papers called AI, the false negative rate the share of the 84 AI papers called human (DOI 10.1515/opis-2022-0158).

The table draws on Table 4 in Walters 2023, as percentages of documents correct, incorrect and uncertain, for the detectors with a card plus the top and bottom of the field (publisher PDF).

Values as printed in Walters 2023, Table 4, DOI 10.1515/opis-2022-0158.

Three detectors stood apart in Walters 2023. Copyleaks and Turnitin were correct on all 126 documents, and Originality.ai was correct on all but two, both human papers that the tool marked uncertain (DOI 10.1515/opis-2022-0158). Among the other 13 detectors in Walters 2023, overall accuracy ran from 63% to 88%.

GPT-4 text separated the field in Walters 2023. Only the top three detectors classified GPT-4 documents with more than 83% accuracy, and the rest tended to call those documents human or uncertain; the average across all 16 was 95% correct on GPT-3.5 papers against 61% on GPT-4 papers in Walters 2023 (publisher PDF). The GPT-4 question, added 9 November 2026, lists the nine cards this record scored on ChatGPT-4 text beside the twelve other GPT-4 rows in the index.

On errors, Walters 2023 found that four detectors each had false positive rates of 10% to 14%, while every other detector had no more than a single false positive among the 42 human papers (publisher page). Crossplag, GPT Radar and Writer had false negative rates of 23% to 36% in Walters 2023; the others reached at most 17%, with a mean of 6.5%. Paid detectors averaged 87% accuracy against 77% for free ones in Walters 2023, and detectors that appear on popular recommendation lists scored 81% against 77% for the rest.

The Walters 2023 author flags Turnitin's training data first: the human papers were student submissions that could have been used to train the Turnitin detector, so Turnitin may be especially accurate on this particular set. All 126 documents in Walters 2023 were undergraduate literature reviews, which the author says may not generalise to other kinds of writing (DOI 10.1515/opis-2022-0158).

The author calls the largest limitation the absence of any paraphrased or edited AI text; every AI paper was submitted unaltered. One coder classified every response, and the human and AI topics were not identical.

Fourteen of the 16 detectors in Walters 2023 have a card in this index. Four joined on 10 November 2026 with the eleventh wave: GPT Radar (76% of 126 documents correct, 36% of the 84 AI papers called human, the highest miss rate of the 16 tied with Writer), Grammica (86% of 126 correct, 0 of 42 human papers called AI), IvyPanda (77% of 126 correct with no incorrect verdict and 23% uncertain) and SEO.ai (72% of 126 correct, 12% of 42 human papers called AI, the third-highest false positive rate), each from Table 4. The Content Detector AI card, for the tool tested here as ContentDetector.ai, was added on 10 November 2026 with the ninth wave of cards; the Writer card was added on 9 November 2026 for a detector withdrawn on 22 December 2025, and the BrandWell card, for the tool tested here as Content at Scale, with the fourth wave of cards the same evening. Each card states the overall, GPT-4 and false positive figures with their bases (DOI 10.1515/opis-2022-0158).

All fourteen appear in the ranking of detectors by independent test record. Table 4 prints Content Detector AI's overall, GPT-3.5 and GPT-4 shares and no separate human-paper count, so its card carries no false positive rate from this record; the same rule leaves the GPT Radar, Grammica and IvyPanda cards without a false positive value, because this record names a human-paper figure only for the eight detectors in the false positive question below, and SEO.ai's card carries the 12% that it names.

Which AI detectors were in Walters 2023? Sixteen, as Table 3 lists them: Content at Scale (now BrandWell), ContentDetector.ai, Copyleaks, Crossplag, Grammica, GPT Radar, GPTZero, IvyPanda, OpenAI's classifier, Originality.ai, Sapling, Scribbr, SEO.ai, Turnitin, Writer and ZeroGPT. Fourteen of the 16 have an index card on 10 November 2026, counted from the table above; Crossplag has none, and the OpenAI classifier is recorded as a withdrawn vendor tool in its own record. Three of the 14 carded tools have since been withdrawn: Writer (22 December 2025), GPT Radar (a redirect to neuraltext.com from 14 August 2026) and Grammica (removed between 15 December 2025 and 12 January 2026), each dated on its card; the Content at Scale and ContentDetector.ai pages have redirected to a Moxby marketplace listing since August 2026, a move rather than a withdrawal, and the OpenAI classifier was withdrawn on 20 July 2023. The shut-down question lists each with its source. How this record is sourced explains why a perfect score on one test is a result, not a verdict, and the home page shows which AI detectors exist and who has tested them.

Related records: Weber-Wulff 2023 tested 14 detectors three months earlier and reached a far lower ceiling, and Elkhatat 2023 found the same GPT-4 gap on a small engineering sample (DOI 10.1515/opis-2022-0158).

Comparisons that read this record first: Scribbr vs Turnitin, since this is the one study that scored both (88% against 100% of 126 documents, 2% against 0 of 42 human papers flagged), and ZeroGPT vs Scribbr vs QuillBot, where it is the one record that tested two of the three free detectors.

Two of the sixteen tools, Sapling and GPTZero, are set side by side on the study's human-paper figures (0 and 2 of 42) in the Sapling vs GPTZero comparison, read 10 November 2026.