The rise of large language models (LLMs) like ChatGPT has brought incredible advancements but also potential risks, such as the spread of misinformation. One proposed solution is watermarking, a technique to embed invisible signals within AI-generated text. But how detectable are these watermarks in the real world? Researchers have investigated this question, focusing on three main families of watermarking schemes: Red-Green, Fixed-Sampling, and Cache-Augmented. Red-Green schemes work by subtly altering the probabilities of word choices based on a secret key. Fixed-Sampling methods use the key to determine the entire sampling process, potentially limiting the diversity of generated text. Cache-Augmented schemes leverage a memory cache to modify outputs only when a specific context hasn't been seen before. The researchers developed statistical tests to detect these watermarks in a black-box setting, meaning they only needed access to the model's text output, not its internal workings. Surprisingly, their tests revealed that current watermarking schemes are more detectable than previously thought, even with limited queries. They tested their methods on various open-source models and found they could reliably identify the presence and type of watermark. Interestingly, when they applied their tests to popular LLMs like GPT-4, Claude 3, and Gemini 1.0 Pro, they found no strong evidence of watermarks. This suggests that widespread watermark deployment in large-scale models is still a work in progress. These findings have significant implications for the future of AI text detection. While watermarking remains a promising approach, the research highlights the need for more robust and less detectable methods. The challenge lies in balancing the need for effective watermarking with the desire to maintain the naturalness and diversity of AI-generated text. Future research might explore entirely new watermarking techniques or focus on strengthening existing ones against detection. The cat-and-mouse game between watermarking and detection will likely continue as AI technology evolves.