In the rapidly evolving world of artificial intelligence, the rise of Large Language Models (LLMs) has brought about both remarkable advancements and potential risks. While LLMs power innovative applications, concerns about misuse and copyright infringement have led to the development of watermarking techniques. These techniques embed hidden signals within the LLM's generated text, allowing for identification of AI-authored content. But what if users could detect these watermarks without any inside knowledge? New research explores this intriguing question, examining whether watermarked LLMs can be identified through cleverly crafted prompts. The study introduces "Water-Probe," a unified method for detecting various watermarking techniques in LLMs. The core idea is that watermarked LLMs exhibit consistent biases under the same watermark key. By designing prompts that trigger repeated sampling with these keys, researchers can analyze the resulting output distributions. Consistent differences across prompts suggest the presence of a watermark. The results are striking: Water-Probe effectively identifies a range of watermarking algorithms, even those claiming to be "distortion-free." This raises important questions about the transparency and security of watermarked LLMs. If users can easily detect watermarks, it could impact trust and potentially open doors to watermark circumvention. The research also delves into strategies for enhancing watermark imperceptibility. One promising approach is the "Water-Bag" strategy, which combines multiple watermark keys to increase randomness. This makes it harder for users to craft prompts that expose the watermark. However, increasing randomness can also decrease watermark robustness, creating a trade-off between detectability and imperceptibility. This research highlights the ongoing cat-and-mouse game between watermarking techniques and detection methods, a dynamic that will continue to shape the future of AI and content creation.