Clickbait headlines continue to challenge the digital media landscape by prioritizing user engagement over content reliability. This study conducts a thorough comparison of methodologies for identifying clickbait, spanning traditional machine learning models, advanced deep learning architectures, and fine-tuned large language models (LLMs). The approaches evaluated include classical classifiers such as Support Vector Machines and Logistic Regression, sequential models like LSTM and BiLSTM, as well as transformer-based models such as BERT and XLM-RoBERTa. Furthermore, the research explores the effectiveness of prompting strategies, including zero-shot and few-shot techniques. Utilizing the Webis Clickbait Corpus 2017—a dataset with significant class imbalance—the experiments demonstrate that fine-tuned XLM-RoBERTa yields the highest performance, achieving a Macro F1 score of 0.8169, which showcases its ability to capture subtle semantic patterns in imbalanced data. Although prompting techniques offer computational efficiency, they fall short in delivering competitive results. This study highlights the importance of employing fine-tuned LLMs for reliable clickbait detection while emphasizing the critical role of evaluation metrics, such as Macro F1, for ensuring balanced performance across all classes in datasets with uneven distributions.