Large language models (LLMs) have become part of everyday life, and we increasingly rely on their seemingly exceptional performance. But are they truly this powerful when it comes to concrete applications such as linguistic analysis? This talk explores the capabilities and limitations of LLMs in the context of empirical linguistic research. Beginning with an overview of their architecture and underlying principles, the discussion will examine both the strengths of LLMs, such as their performance in language understanding and generation tasks, and their shortcomings, including bias, opacity, and hallucination. The second part of the talk investigates how LLMs perform on concrete linguistic tasks, providing a critical assessment of their performance in real-world analysis.