Study finds Pangram beats GPTZero, Turnitin at spotting AI text

The paper tested four of the main AI text detectors used in universities on 160 documents of known provenance. Pangram outperformed GPTZero, Copyleaks, and Turnitin, especially on hybrid texts and on AI-written passages rewritten through a 'humanising' prompt designed to resemble student behaviour. All four tools correctly identified fully human-written documents.

Published in the International Journal for Educational Integrity by Marijke Van Vlasselaer, Filip Van Droogenbroeck and Bram Spruyt of the Vrije Universiteit Brussel, the study then applied Pangram alone to 1,163 master's theses submitted in academic year 2024-2025. 45.5% of them were flagged, 'typically indicating low to moderate levels of AI-associated text'.

The abstract reports that the other three detectors 'significantly underestimated GenAI content, particularly for texts generated with the most advanced model'. It does not name the model. 'False positives were rare across all tools,' the authors write, 'suggesting improvement compared to earlier studies.' Two researchers we follow posted the paper on publication.

The authors close with a caution: detection tools 'can provide useful initial flags' but 'should not be used as sole evidence in high-stakes decision-making but should be implemented in a broader evaluation strategy'.