Generative artificial intelligence (GenAI) is rapidly changing how digital content is created and consumed. Two widely used GenAI technologies are deepfakes and large language models (LLMs). Deepfakes generate or modify images, videos, audio, and text that imitate real people, while LLMs provide robust language understanding, reasoning, and multimodal coordination. When combined, these technologies significantly increase the realism, speed, and accessibility of synthetic media, raising concerns about misinformation, impersonation, and loss of digital trust. At the same time, the same reasoning capabilities that enable deepfake generation can also be leveraged for detection, verification, and mitigation. This article explores how LLMs strengthen deepfake generation by enabling realistic scripts, coordinated multimodal outputs, and scalable automation. Furthermore, it highlights how LLMs can also be used to fight deepfakes through semantic analysis, cross-modal verification, and provenance-based safeguards. By examining this dual role in an agentic AI setting, the article emphasizes why LLMs are