Detecting bots on social media platforms is a major challenge, as these automated entities are constantly evolving to evade detection. In this study, we investigate the main features that contribute to the difficulty of bot detection. Leveraging the TwiBot-20 dataset, we analyze the characteristics of misclassified accounts and explore the reasons behind their erroneous classification. Our approach combines feature engineering, Machine Learning with Random Forest, and the interpretation of model predictions using SHAP (SHapley Additive exPlanations) values. We employ clustering techniques to identify patterns in feature contributions and provide insights into the complexities of distinguishing between human and automated accounts. Our findings highlight the nature of bot detection and the need for advanced methods to address the problem of social media manipulation.