<![CDATA[The rapid advancement of Generative AI driven by Large Language Models (LLMs) has raised significant concerns regarding the integrity of digital information. Although text detection systems are evolving rapidly, the majority of current instruments predominantly support English, thereby leaving a substantial research gap in low-resource languages such as Indonesian. This study aims to conduct a comparative performance analysis between XGBoost and IndoBERT for detecting Indonesian text generated by Generative AI. Experiments were executed using a structured, hand-crafted dataset spanning three main domains: Science and Technology, Social and Culture, and Politics. The XGBoost model was evaluated utilizing Term Frequency-Inverse Document Frequency (TF-IDF) feature extraction, whereas IndoBERT relied on bidirectional contextual semantic understanding. The evaluation results on the controlled dataset demonstrated that both architectures achieved perfect accuracy (100%). However, under a stress test employing adversarial trap sentences designed with keyword stuffing techniques, the performance of the two models exhibited a sharp contrast. The global accuracy of XGBoost plummeted drastically to 50.00% because the TF-IDF representation failed to recognize the core logical essence of the sentences and was easily misled by the dominance of individual statistical keywords. Conversely, IndoBERT proved to be more robust by maintaining a global accuracy of 66.67% and recording maximum Precision, Recall, and F1-Score (1.000000) within the Social and Culture class. This demonstrates the capability of the Self-Attention Mechanism in capturing latent context and complex linguistic structures. The implications of these findings emphasize that in countering advanced text manipulation (covert hoaxes), the deployment of context-based models like IndoBERT is far more reliable. Nevertheless, architectural selection must remain dynamically aligned with local computing infrastructure capacity.]]>
Copyrights © 2026
