Large language models (LLMs) like ChatGPT are impressive, but they can be misused. Think fake news or plagiarism—big problems. Researchers are working on ways to track the output of these models, a bit like a digital signature for text. This is called watermarking. Ideally, a watermark is invisible to the reader, easily detectable by software, and tough enough to survive editing. The challenge? Balancing these three qualities—imperceptibility, efficacy, and robustness—is tricky. A new research paper introduces "WaterPool," a clever approach to LLM watermarking that tackles this balancing act. Existing methods often struggle because the way they generate the watermark clashes with how they detect it. WaterPool uses a two-part system: a "key" to create the watermark and a "mark" to embed it in the text. The key innovation is how WaterPool manages these keys. It uses a vast pool of keys for generation, ensuring the watermark is virtually undetectable. For detection, it employs a semantic search, essentially finding the closest matching key based on the meaning of the text. This makes detection accurate even if the text has been altered. Tests show WaterPool significantly improves existing watermarking techniques, boosting their robustness and detection rates while keeping the watermark invisible. This is a big step towards responsible AI development, making it easier to identify misuse and protect the integrity of information.