This paper investigates whether Large Language Models (LLMs) can serve as zero-shot human activity recognizers (HAR) by analyzing raw IMU data. The study, named HARGPT, demonstrates that LLMs can comprehend raw IMU data and perform HAR tasks without prior training or domain-specific knowledge. HARGPT uses role-play and "think step-by-step" prompting strategies to guide LLMs in interpreting raw sensor data. The system is benchmarked on two public datasets with varying levels of inter-class similarity. Results show that LLMs consistently outperform traditional machine learning and deep learning models in both datasets, achieving an average accuracy of 80% on unseen data. The study highlights that LLMs can process raw sensor data effectively, leveraging their knowledge base to interpret physical world data. The findings suggest that LLMs have significant potential for analyzing raw sensor data in Cyber-Physical Systems (CPS). The paper also discusses the special properties of LLMs, such as their logical reasoning ability and the challenges in generating perfunctory answers. The study concludes that LLMs can be used as a foundational model for HAR with high accuracy and robustness, offering transformative potential for CPS. Future work should focus on standardizing evaluation processes and further exploring the capabilities and limitations of LLMs in real-world applications.This paper investigates whether Large Language Models (LLMs) can serve as zero-shot human activity recognizers (HAR) by analyzing raw IMU data. The study, named HARGPT, demonstrates that LLMs can comprehend raw IMU data and perform HAR tasks without prior training or domain-specific knowledge. HARGPT uses role-play and "think step-by-step" prompting strategies to guide LLMs in interpreting raw sensor data. The system is benchmarked on two public datasets with varying levels of inter-class similarity. Results show that LLMs consistently outperform traditional machine learning and deep learning models in both datasets, achieving an average accuracy of 80% on unseen data. The study highlights that LLMs can process raw sensor data effectively, leveraging their knowledge base to interpret physical world data. The findings suggest that LLMs have significant potential for analyzing raw sensor data in Cyber-Physical Systems (CPS). The paper also discusses the special properties of LLMs, such as their logical reasoning ability and the challenges in generating perfunctory answers. The study concludes that LLMs can be used as a foundational model for HAR with high accuracy and robustness, offering transformative potential for CPS. Future work should focus on standardizing evaluation processes and further exploring the capabilities and limitations of LLMs in real-world applications.