Detecting the Undetectable: A Critical Synthesis of Machine Learning Applications for Academic Dishonesty Identification in Online Learning Environments

Abstract

The rapid and sustained shift toward online learning has rendered traditional academic integrity mechanisms increasingly
inadequate, creating urgent demand for scalable, automated detection approaches. This paper critically examines the application
of machine learning models for identifying academic dishonesty in online learning environments, synthesizing peer-reviewed
literature from 2018 to 2025 across computer science, educational psychology, and learning analytics. Through a systematic
qualitative synthesis of 48 empirical studies, we identify three dominant detection paradigms: behavioral pattern analysis, textual
stylometry, and multimodal assessment forensics. The analysis reveals a fundamental tension between detection accuracy and
educational equity, as models achieving high classification performance disproportionately flag submissions from non native
English speakers, students with disabilities, and those accessing assessment from unstable technological environments. We propose
a tripartite theoretical framework—the Detection Equity Trilemma—that conceptualizes the necessary trade offs among sensitivity,
specificity, and demographic parity. Findings indicate that current machine learning approaches, while technically sophisticated,
embed normative assumptions about legitimate student behavior that systematically disadvantage already marginalized learners.
The paper concludes with recommendations for validity focused model development, transparency mandates, and the integration
of human review protocols that prioritize formative integrity over punitive surveillance.