In this study, we focus on competitive Arabic debates and employ language models to classify debaters’ persuasion modes. Specifically, we aim to identify and categorize ethos, pathos, and logos-the foundational rhetorical strategies underpinning persuasive discourse. We propose two distinct approaches for persuasion mode classification to achieve these objectives. Firstly, we train a baseline language model using the ULMFiT method with an LSTM model. Secondly, we explore the fine-tuning of Language Models pre-trained on Arabic texts to leverage the contextual knowledge captured by pre-trained models. In the early stages of this research, we present preliminary results from each approach, outlining their respective strengths and limitations. Our findings demonstrate that CAMeLBERT outperforms other language models in the accuracy of classification and fairness. The CAMeLBERT reaches 91.7% accuracy, 90% f1-score, and less than 3.6% and 1.4% in demographic parity and equal opportunity differences, respectively.