DistilBERT-Based Detection of AI-Generated Text in Online Assessments: Ethical and Pedagogical Implications

Abstract

The rapid shift toward online and distance learning has positioned digital assessment as a cornerstone of higher education while also challenging academic integrity due to the accessibility of generative artificial intelligence (GenAI). While technical research into AI detection is expanding, there remains a critical gap in understanding how detection outcomes can be ethically and pedagogically integrated into digital learning environments. This study evaluates a fine-tuned DistilBERT-based model for detecting AI-generated text, situating its technical performance within a learning-centered framework. Using a large-scale dataset of over 28,000 human-written and AI-generated essays, the model demonstrated exceptional robustness, achieving an overall accuracy of 99%, an AUC of 0.9999, and balanced F1 scores of 0.99. Beyond technical metrics, this research redefines AI detection by shifting the narrative from a punitive, surveillance-oriented mechanism to a supportive learning analytics tool. By interpreting detection results alongside instructional indicators, the study demonstrates how these technologies can inform assessment redesign, enhance transparency, and foster learner trust. The findings contribute to the field of digital education by providing a roadmap for the responsible integration of AI detection into assessment ecosystems, ensuring that technological precision serves the broader goals of fairness and pedagogical innovation.