Van Vlasselaer et al. 2026: who wrote this? GPTZero, Pangram, Copyleaks and Turnitin on 160 academic papers

This record joins the studies index as the first preregistered detector test indexed here. A correction published on 6 October 2026 renamed the generator label from "GPT-4o-Deep-Research" to "ChatGPT Deep Research mode"; the authors state that the data, analyses and conclusions are unchanged (DOI 10.1007/s40979-026-00236-8, read 9 November 2026). Every figure on this record was checked against the published article on 9 November 2026.

Van Vlasselaer 2026 tested four detectors, all of which have cards in this index (DOI 10.1007/s40979-026-00226-w, Methodology, read 9 November 2026).

Detectors as named in Van Vlasselaer et al. 2026, DOI 10.1007/s40979-026-00226-w.

The synthetic set holds 160 English-language papers of at least 4,000 words in four categories of 40. The human papers were written by non-native English-speaking graduate students at the authors' faculty before 2019. The fully AI papers were generated in May 2025 with ChatGPT Deep Research mode, running from 4,000 to 12,000 words with a mean of 6,785. The hybrid papers replaced the literature review and conclusion of each human paper with GPT-4o text, for a median ground truth of 44.6% AI; the humanised papers rewrote those AI sections with a GPT-4o prompt asking for a student-like voice, for a median of 44.0% AI (DOI 10.1007/s40979-026-00226-w, Data collection and method, read 9 November 2026).

Sample as described in Van Vlasselaer et al. 2026, DOI 10.1007/s40979-026-00226-w, Table 2.

Scores were classified on the scale Weber-Wulff et al. 2023 used: for a fully AI paper a score of 80% or more is a true positive and under 20% a false negative, with partial bands between; for a human paper a score under 20% is a true negative; for hybrid and humanised papers a true positive is a score inside that paper's known range of AI content. Turnitin reports an asterisk instead of a number for scores under 20%, and those cases were treated as the lowest band. The study was preregistered on AsPredicted (#228383); the humanised category and per-category reporting were added afterwards, and the authors say so (DOI 10.1007/s40979-026-00226-w, Classification accuracy and error analysis, read 9 November 2026).

Tables 4 and 5 of Van Vlasselaer 2026 give classification results per category; the Results prose gives the same figures (DOI 10.1007/s40979-026-00226-w, read 9 November 2026).

Values as printed in Van Vlasselaer et al. 2026, DOI 10.1007/s40979-026-00226-w, Tables 4, 5 and 6 and Results.

On the fully AI papers the authors write that "Turnitin, GPTZero, and Copyleaks completely failed to detect the AI-generated content"; the Figure 1 note records that 19 of the 40 fully AI papers, 7 hybrid, 6 humanised and 3 human papers came back from Turnitin with an asterisk rather than a score (DOI 10.1007/s40979-026-00226-w, Descriptive results and Figure 1, read 9 November 2026). The authors attribute the gap between the fully AI and mixed categories to the generator: Deep Research mode against GPT-4o.

The descriptive results say GPTZero "showed a small false positive rate" on human papers, which the mixed model puts at a mean overestimate of 1.95 points, while under the 20% rule every human paper from every tool counted as a true negative (DOI 10.1007/s40979-026-00226-w, Descriptive results and Inferential analysis, read 9 November 2026). On the mixed categories Pangram's average error was not significantly different from zero, while Turnitin, Copyleaks and GPTZero underestimated AI content by 10 to 24 points.

In the second part of the study Pangram alone was run on 1,163 master's theses from one faculty in the 2024 to 2025 academic year. It flagged 529 (45.5%); among flagged theses the mean estimated AI share was 34.0% and the interquartile range 17.0% to 49.0% (DOI 10.1007/s40979-026-00226-w, Real-world application and Table 6, read 9 November 2026). The authors state that with no ground truth these percentages cannot be read as confirmed prevalence of AI use.

The authors call the findings indicative rather than exhaustive: 160 papers, four tools and one generator for the fully AI category, valid for a snapshot in time. Only one humanisation strategy was tested, a single GPT-4o prompt, so detection of student-edited hybrid text may be lower than shown, and all human papers came from one faculty (DOI 10.1007/s40979-026-00226-w, Limitations, read 9 November 2026). The Discussion adds that some tools may set conservative thresholds to avoid flagging human text, which would raise false negatives, and recommends that a flag from any tool trigger review rather than a conclusion.

All four detectors in Van Vlasselaer 2026 have cards in this index. Each card records the fully AI result and the hybrid and humanised accuracies on 40 papers per category (DOI 10.1007/s40979-026-00226-w, read 9 November 2026).

All four appear in the ranking of detectors by independent test record, and the zero false positive count on 40 human papers is noted on the false positive rate ranking with its base. How this record is sourced explains the base rule, and the home page lists which AI detectors exist and who has tested them.

Related records: Weber-Wulff 2023 supplied the classification scale; Jabarian and Imas 2025 tested Pangram and GPTZero on 1,992 passages; the Turnitin accuracy record carries the vendor's own statements (DOI 10.1007/s40979-026-00226-w, read 9 November 2026).