The rapid advancement of generative AI models, particularly Multimodal Large Language Models capable of producing both textual and visual content, has revealed inherent biases embedded within these systems. These biases—spanning dimensions such as gender, race, socioeconomic status, and cultural representation—reflect and amplify societal inequalities present in training data. This study systematically examines these biases by analyzing responses to prompts in both text-to-image and text generation tasks. By comparing biases observed in visual and textual outputs across diverse social categories, we uncover patterns of stereotyping, exclusion, and demographic misrepresentation. Statistical and qualitative evaluations highlight the consistency and variability of biases across modalities and models, revealing potential risks to trust, fairness, and societal stability in high-stakes AI applications. By addressing these challenges, we contribute to the development of safe, fair, and inclusive AI systems, fostering ethical practices in their design and deployment.