The widespread use of social media platforms has led to a surge in malicious automated accounts, commonly known as bots, which pose significant threats by spreading misinformation, manipulating public sentiment, amplifying propaganda and engaging in spam activities. Detecting these bots effectively is critical to maintaining the authenticity and safety of digital communication environments. This paper presents a comparative study between traditional machine learning approaches and graph neural network (GNN)-based models in the task of social media bot detection. Traditional methods typically rely heavily on handcrafted features derived from user behaviour, textual content, and account metadata, often requiring domain expertise, are sensitive to feature selection and may struggle with generalization across platforms. In contrast, GNNs leverage the inherent structure of social networks, capturing relational and topological patterns often indicative of coordinated inauthentic behaviour. We evaluate both paradigms using publicly available datasets, considering metrics such as accuracy, F1 score, robustness to adversarial behaviour, and computational efficiency. Our results show that while traditional models can achieve competitive performance with well-engineered features, GNNs offer superior adaptability by learning from the underlying graph structure without the need for extensive manual feature engineering. Additionally, we explore the trade-offs between model interpretability, scalability, and detection efficacy. This study contributes to a deeper understanding of the strengths and limitations of both approaches, offering practical guidance for researchers and practitioners in designing more resilient bot detection systems.