<![CDATA[The rise of false information in Indonesia has turned into a pressing problem in today's digital era, primarily driven by the heavy reliance on social media as the dominant source of information. This study aims to evaluate the effectiveness of the Transformer model in detecting fake news written in Indonesian and to compare its performance with traditional methods such as LSTM and Naïve Bayes. An experimental quantitative approach was employed, utilizing a curated dataset of verified real and fake news articles. The data underwent preprocessing stages, including text cleaning, tokenization, and transformation into numerical vector formats prior to model training. The results demonstrate that the Transformer model surpasses other methods, achieving an accuracy of 92.6% and outperforming both LSTM and Naïve Bayes in all principal evaluation metrics. In addition, the Transformer model efficiently detects typical linguistic patterns found in hoax content, including exaggerated expressions, conspiracy-related terms, and repetitive sentence constructions. Validation on an external dataset further confirms the model’s ability to maintain performance stability beyond the initial training data. Despite its promising results, the implementation of this model faces several challenges, including the limited availability of Indonesian-language datasets and concerns related to data ethics and privacy. This study contributes theoretically to the advancement of Transformer-based NLP and practically supports the enhancement of digital literacy and the development of contextual and adaptive hoax detection systems in Indonesia.]]>
