Comparative accuracy of AI-based plagiarism detection tools: an enhanced systematic review

Abstract

Turnitin AI has detected machine-generated text with accuracy rates ranging from 92% to 100% and approximately 5.3% false negative rate (Gosling et al., 2024; Díaz Arce, 2023). OriginalityAI has achieved near-perfect accuracy (98%-100%), while Sapling has demonstrated 97% accuracy (Akram, 2024; Howard et al., 2024). In contrast, human evaluators have achieved accuracy scores ranging from 53% to 79.41% (Dawson et al., 2019). Various studies have documented these tools' strong performance across text types including academic papers, scientific abstracts, and personal statements, though few studies have compared results across academic disciplines. For instance, one study documented high performance in computer science, physics, and mathematics (Akram, 2024), while another observed variations in GPTZero's performance between biology and computer assignments (Steponenaite & Barakat, 2023). No studies have reported results for Winston AI.