The phenomenon of fake news disseminates fabricated information presented in a news-like fashion, posing significant challenges for news agencies regarding accurate processing and verification. The dissemination of fake material could incite or defame prominent entities or individuals and may even serve the personal agendas of its makers, thereby posing societal encounters. Differentiating between fake and real news poses a substantial problem, mostly stemming from the constraints imposed by limited topic knowledge and time limitations. Based on the survey findings, Banten, DKI Jakarta, and West Java are the evident regions with the highest exposure to hoaxes and misinformation among their populations. An artificial intelligence (AI) methodology, the transformers, employs natural language processing (NLP), leveraging deep learning architectures to mitigate fake news. Transformers use a robust attention mechanism to concurrently process textual data and generate comprehensive and contextually informed word representations. A prior investigation demonstrates the higher performance of BERT, a transformer-based model, compared to the non-transformer approach. However, several studies have indicated that the performance of BERT models is potentially enhanced by utilizing advanced variants such as ALBERT and RoBERTa. Thus, further investigation is necessary to improve the utilization of modified BERT models in detecting fabricated news in Bahasa Indonesia. This study investigates various transformer models and clarifies that ALBERT performs superior to other models, achieving an accuracy of 87.6%, precision of 86.9%, F1-score of 86.9%, and a run-time of 174.5 seconds per epoch. The source code can be accessed at the following GitHub repository: github.com/Shafna81/fakenewsdetection.git.