Socially interactive agents (SIAs) simulate essential aspects of human conversation, encompassing both verbal and non-verbal behaviors, and are increasingly integrated into diverse sectors such as healthcare and education. Accurately interpreting and generating non-verbal cues is crucial for enhancing communication effectiveness and user satisfaction. However, the reliance of current research on data-driven approaches in behavior generation for SIAs often results in models inheriting biases from biased real-world datasets, potentially reinforcing societal stereotypes and compromising the ethical integrity of these agents. In this paper, we focus on identifying gender biases in generative models of facial non-verbal behaviors, including gaze, head movements, and facial expressions. By analyzing both real-world interaction data and generated data from a state-of-the-art generative model, and employing a gender classifier, we aim to highlight gender biases present in both types of datasets. The findings from this research initiate discussions on strategies to analyze and mitigate these biases, thereby promoting the development of more inclusive and fair SIAs.