False Positives in AI Writing Detection: A Small-Scale Empirical Study Using Authentic Filipino Student Essays

This research note reports a small-scale exploratory study into how two AI writing detectors, ZeroGPT and Copyleaks, classify authentic student essays. A total of ten anonymised college-student essays were analysed to observe misclassification patterns, particularly false positives, where human-written content has been incorrectly flagged as AI-generated. The assessment was conducted focusing on essays produced by Filipino undergraduates whose nonnative English writing may have features that could lead to misclassification by the detectors. Results show that both detectors inconsistently classified the same set of essays: five essays were labelled as “AI-generated,” while the other five were labelled as “Human,” when in fact all the texts were authentically written by students. These results point to the potential misclassification risks when AI detection tools are used within educational contexts, where it is commonplace for teachers to make important decisions about academic integrity based on the outputs of detectors. The present study underlines the need to validate the outputs of AI detectors with human judgment and advises educators against the use of these tools as sole evidence of misconduct. Implications for ICT-supported assessment practices and policies of academic honesty are discussed, together with recommendations for a more responsible integration of AI detection tools in educational contexts. The results described above have implications for the practice of academic integrity in all countries, but also specifically for those countries with multilingual student bodies whose writing styles may differ from the writing style used to train the detectors.
https://doi.org/10.64233/VYVI9613