Falls among the elderly pose significant health risks and require timely detection to prevent serious consequences. This paper investigates the preliminary application of vision language models (VLMs) for fall detection using static RGB and depth images. The study addresses the problem of detecting fall scenarios from still images by prompting VLMs to evaluate posture cues. Using the Fall Detection Dataset (FDD), we evaluated two state-of-the-art models, Gemma-3-4B-IT and LLaMA-3.2-11B Vision-Instruct, and examined their ability to identify fall-related patterns. Although limited by the static nature of the input, our results reveal that VLMs can effectively detect fall scenarios, offering interpretable output that could support healthcare professionals. The study also discusses potential limitations, including computational demands and reduced accuracy compared to specialized models. These findings contribute to the foundation for future multimodal, real-time, and explainable fall detection systems in smart healthcare environments.