Background: The dissemination of misinformation through online platforms is a serious concern in today’s information-based world. As it has become easier to publish and share information online, fake news has emerged as a critical threat, influencing public perception and distorting facts. It is therefore important to correctly identify fake news in order to maintain the integrity of information and ensure public awareness.

Objectives: The objective of this study is to perform a comparative analysis ofthe Deep Learning (DL) models namely Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional Encoder Representations from Transformers (BERT) for fake news detection. By evaluating the performance of these models using metrics such as accuracy, precision, recall, and F1-score, the study aims toidentify the most effective method for detecting misinformation in text data.

Methods: The research adopts an experimental approach, by training and testing various DL models on a labeled fake news dataset sourced from Kaggle. Preprocessing steps such as tokenization and sequence padding are applied to prepare the text data for model input. The LSTM and GRU models are implemented using recurrent neural network layers, while BERT is employed using transfer learning techniques. Each model is evaluated on the same dataset to ensure a fair and consistent comparison.

Results: The evaluation indicates that transformer-based models perform significantly better than recurrent neural networks (RNNs) in fake news detection. The BERT model achieved the highest accuracy, reaching 99% among the evaluated models. The LSTM and GRU models achieved approximately 98% and 93% accuracy, respectively. The experimental results highlight the effectiveness of contextual word embeddings and multi-head attention mechanisms in capturing complex textual patterns.

Conclusions: This work demonstrates the effectiveness of the BERT model compared to conventional RNNs for fake news detection. By leveraging contextual understanding of textual data, BERT proves to be a robust tool for detecting misinformation with high accuracy. These results highlight the importance of transformer-based methods in constructing more stable and trustworthy fake news detection systems.