Abstract:
The emergence of social media on the internet has increased the rate at which people spread misinformation and fake news, posing serious societal, political, and economic dangers. This paper created a predictive intelligence model to detect fake news, guided by the objective of developing and evaluating a transformer-based model for classifying news as true or fake while also assessing its performance relative to traditional machine learning baselines. Advanced natural language processing and machine learning methods were used to define news as true or fake by applying the Bidirectional Encoder Representations from Transformers (BERT) model. Training and testing of the model were conducted using a Jupyter Notebook with the assistance of a GPU on Google Colab and a labeled Fake and True news dataset available on the Kaggle repository. Accuracy, precision, and recall were used to evaluate performance, and it was found that BERT outperformed the traditional machine learning baselines, achieving higher accuracy while preserving good precision and recall. In addition to technical performance, the study highlights the need to make detection systems local to the language and recommends the inclusion of Kiswahili, Sheng, and native dialects to make them more inclusive and relevant in the real-life Kenyan setting. The results correspond to existing literature indicating the success of transformer-based models in countering misinformation in dynamic web-based contexts. The impact of this study is both theoretical, through the development of deep learning methods for misinformation detection, and practical, as the study provides a scalable framework for policymakers, social media sites, and fact-checking institutions. Further research on multilingual adaptation, integration with real-time monitoring, and ethical safeguards against careless utilization of AI in the fight against misinformation is recommended.
