Tri Polar Analysis: Comparative Evaluation of Political Bias in Fake News Detection Using Three Identically Structured Machine Learning Models

The rapid spread of fake news and overt political bias in the media has prompted a need for robust detection systems. This study explores the impact of political bias on fake news detection by developing three identical machine learning models. A control model was trained on 40,000 pieces of non-biased news data, while two additional models were trained on 40,000 heavily biased news articles each—one with a left-wing (Democratic) bias and the other with a right-wing (Republican) bias. To assess cross-political performance, 20% of the left-wing data was used to test the right-wing model and vice versa, with each biased model also evaluated on a separate 20% non-biased dataset. Results indicate that the left-wing biased model achieved an accuracy of 73.0% on its primary test set but dropped to 58.0% when evaluated on right-wing data. Conversely, the right-wing biased model recorded 74.7% accuracy on its primary test set and 80.2% on left-wing data. In contrast, the control model attained a consistently high accuracy of 89.0% across both politically opposed and unbiased test sets. Precision and recall metrics further support these findings: the control model achieved a precision of 90.1% and a recall of 88.5%, outperforming the left-wing model (precision 87.2%, recall 85.0%) and the right-wing model (precision 79.0%, recall 77.5%). These outcomes suggest that while politically biased models perform well on data aligned with their training bias, their effectiveness diminishes with opposing data, underscoring the importance of unbiased training data for accurate fake news detection. However, the study was limited by constraints in computing power and the challenges associated with cleaning and sourcing high-quality biased news data, which may impact the overall scalability and generalizability of the approach.