<![CDATA[This study aims to analyze the developments, gaps, and methodological novelties in Indonesian-language hoax detection techniques, focusing on the transition from classical machine learning approaches to deep learning models and transformer-based architectures. The research method used was a non-SLR literature review of 31 Scopus-indexed national, international, and international journal articles published between 2019–2025. Data is extracted through a structured analysis sheet that includes dataset characteristics, algorithm models, linguistic features, evaluation metrics, and quantitative performance results. The results show a significant evolution: SVM-based and Naive Bayes-based approaches produce an accuracy of 85–97%, while deep learning models such as CNN, LSTM, and GRU achieve 90–99%. Transformer-based models such as IndoBERT and RoBERTa record the highest performance with an accuracy of 94–99.5%, along with better computing efficiency. Hybrid approaches such as Bi-LSTM + IndoBERT and CNN-LSTM show the most stable and superior performance compared to single models. The study also identified key gaps in model interpretability, unbalanced data handling, and integration of multimodal approaches. In conclusion, the development of the Indonesian hoax detection methodology shows a direction towards a more contextual, efficient, and adaptive system. These findings make an important conceptual contribution to the development of a hoax detection system that is more transparent, reliable, and relevant to today's digital information challenges.]]>
