Over the last years, with the explosive growth of social media, huge amounts of rumors have been rapidly spread on the internet. Indeed, the proliferation of malicious misinformation and nasty rumors in social media can have harmful effects on individuals and society. In this paper, we investigate the content of the fake news in the Arabic world through the information posted on YouTube. Our contribution is threefold. First, we introduce a novel Arab corpus for the task of fake news analysis, covering the topics most concerned by rumors. We describe the corpus and the data collection process in detail. Second, we present several exploratory analysis on the harvested data in order to retrieve some useful knowledge about the transmission of rumors for the studied topics. Third, we test the possibility of discrimination between rumor and no rumor comments using three machine learning classifiers namely, Support Vector Machine (SVM), Decision Tree (DT) and Multinomial Naïve Bayes (MNB).