False Positives in Artificial Intelligence Detectors and Their Impact on Academic Quality Assurance: A Systematic Review

Generative artificial intelligence has transformed assessment in higher education, accelerating the adoption of automated text classifiers to uphold academic integrity; nonetheless, their diagnostic reliability remains contested due to systemic algorithmic biases. This systematic review synthesizes empirical evidence on the incidence of false positives in AI-generated text detectors, their algorithmic and sociolinguistic drivers, and their broader ramifications for institutional quality assurance. Following the PRISMA 2020 guidelines and eligibility criteria based on the SPICE framework, searches were conducted in Web of Science, Scopus, and SciELO, resulting in a qualitative and narrative synthesis of 30 primary empirical studies; the protocol was registered in OSF. Preliminary results from the studies indicate that current detectors rely heavily on reductionist probabilistic metrics, such as perplexity and burstiness, which frequently misclassify structured academic discourse and non-native English writing as synthetic text. Furthermore, the synthesized evidence points to the possibility that punitive enforcement may disproportionately disadvantage specific linguistic cohorts, potentially exacerbating assessment anxiety and compromising evaluative equity. The review suggests that relying exclusively on automated AI detection represents a documented institutional vulnerability for assessment fairness and academic quality assurance, particularly within English-medium higher education contexts where these tools are most extensively implemented.