On hybrid and humanised AI papers, Pangram correctly identified AI content in 37 of 40 cases while GPTZero and Copyleaks performed poorly and Turnitin declined on humanised texts¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Pangram correctly identified AI content within the expected range in 37 of 40 cases for both hybrid and humanised categories (strict accuracy 92.5%; inclusive 95.0%). → Marijke Van Vlasselaer 2026 q2 i? GPTZero classified all 40 hybrid papers as false negatives (0.0% accuracy) and achieved 2.5% on humanised papers; Copyleaks achieved 30.0% and 22.5%; Turnitin achieved 60.0% on hybrid but dropped to 50.0% on humanised texts. → Marijke Van Vlasselaer 2026 (2)

Evidence¶

Marijke Van Vlasselaer 2026¶

Marijke Van Vlasselaer, Filip Van Droogenbroeck and Bram Spruyt. (2026). Who wrote this? Evaluating the reliability of AI detection tools in higher education. International Journal for Educational Integrity. https://doi.org/10.1007/s40979-026-00226-w

q2 · i? · design · r2

Classification accuracy analysis of the 40 hybrid and 40 humanised papers in the synthetic dataset, scored against each paper's known ground-truth range of AI content. The article reports Pangram identified "AI content within the expected range in 37 of 40 cases for both categories".

"For Hybrid and Humanised papers, Pangram again outperformed the other tools, cor - rectly identifying AI content within the expected range in 37 of 40 cases for both catego- ries (strict accuracy = 92.5%; inclusive = 95.0%)."

Marijke Van Vlasselaer 2026 (2)¶

Marijke Van Vlasselaer, Filip Van Droogenbroeck and Bram Spruyt. (2026). Who wrote this? Evaluating the reliability of AI detection tools in higher education. International Journal for Educational Integrity. https://doi.org/10.1007/s40979-026-00226-w

q2 · i? · design · r2

Same classification analysis for the comparison tools. The article reports Copyleaks at 30.0% (hybrid) and 22.5% (humanised), and Turnitin at 60.0% declining to 50.0% on humanised texts; no effect sizes are printed.

"For Hybrid papers, GPTZero classified all 40 papers as False Negatives with an accuracy of 0.0%. Its performance on Humanised papers was nearly identical, with a single paper correctly identified (2.5% accuracy)."

Discussion¶

Related Claims¶

• A linear mixed-effects model of detection error showed significant paper type, tool, and interaction effects, with underestimation largest for fully AI-generated papers and Pangram consistently producing the smallest errors — related
• Applied to 1,163 real master's theses, Pangram flagged 529 (45.5%) as containing AI-generated content, mostly at low-to-moderate levels — related
• On fully AI-generated papers produced with ChatGPT Deep Research, Pangram achieved 65% strict accuracy while Turnitin, GPTZero and Copyleaks classified essentially all papers as false negatives — related
• Popular plagiarism detection tools detect plagiarism in AI-assisted essays in less than 15% of cases — related
• After humanization with Undetectable AI, fewer than 4% of AI-labeled rewrites remain detectable (FNR > 96%) for Pangram and GPTZero — reports the opposite
• No detection threshold simultaneously keeps original flag rates, light-edit capture, and AI-pool false negatives low — related