Detecting AI-Generated Text in Educational Content: Leveraging Machine Learning and Explainable AI for Academic Integrity

Detection tools can be effective but have limitations: Machine learning models achieved 83% accuracy in detecting AI-generated text, but this performance dropped significantly when analyzing shorter content, suggesting that educators cannot rely on these tools as foolproof solutions for all student submissions.

AI and human writing have distinct linguistic fingerprints: The study identified that human-written content tends to use practical language (like "use" and "allow"), while AI-generated text favors more abstract and formal terms (like "realm" and "employ")—a finding that could help educators develop intuition for spotting AI-generated work.

Specialized detection models outperform general-purpose tools: The researchers' focused model achieved better accuracy than GPTZero, a widely-used commercial detector, demonstrating that customized tools tailored to specific educational contexts may be more reliable than one-size-fits-all solutions.

Explainable AI adds transparency to detection decisions: By using XAI techniques, the study revealed why the detection model made its predictions rather than treating it as a "black box," which is important for maintaining trust and fairness when flagging student work.

Dataset size and scope matter for real-world application: The study used only 1,000 samples from one AI system (ChatGPT), which may limit how well these findings generalize to other AI tools or writing contexts that educators encounter in practice.