Development of a Text-Based Anti-Cheat Component Using ALBERT for Detecting Cheating in Automated Interviews
Yakobus Iryanto Prasethio, Agung Dewandaru, Gusti Ayu Putri Saptawati · 2024
Maintaining the integrity of online recruitment is increasingly challenging with AI tools like ChatGPT. This paper presents a text-based anti-cheat system using the ALBERT model to detect AI-generated text responses in automated interviews setting. ALBERT was chosen for its efficiency and contextual understanding. The system leverages external indicators such as POS tags, sentiment analysis, readability scores, perplexity, and typing behaviors. Fine-tuned ALBERT and Logistic Regression models were integrated asynchronously into a recruitment system. The ALBERT model achieved 81% accuracy, 87% precision, 79% recall, and 82 % F1-Score, while Logistic Regression Model achieved 78% accuracy, 73% precision, 73% recall, and 72% F1-Score. This result shows that machine learning based classifier can help the identification of cheating behavior.