This paper investigates the potential of Large Language Models (LLMs), such as GPT-4, Cohere, and Gemini, to generate synthetic samples of time-series sensor data using only a few-shot learning (3 samples) without fine-tuning. We aim to highlight their viability in augmenting datasets with minimal data, addressing data scarcity and class imbalance challenges. We evaluate Human Activity Recognition (HAR) tasks from wearable device sensors as a use case in this investigation. Our findings demonstrate that LLMs can produce high-quality synthetic samples in less imbalanced datasets, achieving competitive results compared to traditional generative models. However, the LLM performance decreases with more imbalanced datasets, where the generated synthetic data lacks diversity. We also observed that classification models trained with LLM-generated samples showed more stability in terms of confidence intervals, with the Gemini model consistently producing more stable data. We also present a framework for evaluating synthetic data generation methods, showing the trade-off between synthetic and real-world data and suggesting practical directions for future work addressing data scarcity and balance limitations.