ARE AI DETECTORS FAIR TO EFL LEARNERS? A LIBRARY RESEARCH ON DETECTION BIAS AND ACADEMIC INTEGRITY

Abstract

AI detection tools such as Turnitin AI Detection, GPTZero, and Copyleaks are now embedded in university assessment workflows to identify machine generated text, yet their performance on writing produced by non-native English speakers remains poorly understood. This critical library research synthesizes peer-reviewed literature published between 2022 and 2026 to examine three interrelated questions: how AI detectors perform on non-native English writing, what linguistic mechanisms produce detection bias against EFL learners, and what pedagogical and institutional implications follow from deploying these tools in linguistically diverse contexts. Literature was identified through Consensus AI using six targeted queries with filters restricting results to empirical, peer-reviewed studies. The review reveals that false positive rates for non-native writers exceed 61% in controlled studies, a disparity rooted in the overlap between the statistical profile of EFL writing and the features that detectors associate with machine generated text. Lexical predictability and syntactic uniformity, which are developmental characteristics of second language writing, produce the same low perplexity and uniform burstiness that detectors treat as markers of AI authorship. The uncritical adoption of these tools risks generating unjust accusations that compound existing inequities for EFL learners. Institutions should supplement automated detection with informed human review, redesign assessment toward process-oriented approaches, and develop context-sensitive policies that account for the linguistic profiles of second language writers.