Current fake news detection models often use distinct datasets in their respective train/test processes. This thesis focuses on textual detection models and explores whether classification performance can be affected by in- and between-dataset biases. The result confirms that biases exist and impact the efficacy of detection models.