The rise of sophisticated AI text generators like ChatGPT has opened exciting possibilities but also raised concerns about misuse. One proposed solution is *watermarking*, a technique to embed hidden signals within AI-generated text, allowing its origin to be verified. But how robust are these watermarks against manipulation? Researchers have developed 'WaterPark,' a comprehensive platform to test the resilience of AI watermarks against various attacks. Their findings reveal surprising vulnerabilities, challenging previous assumptions about the security of these digital fingerprints. While some watermarks hold up well against simple alterations like changing word order or introducing typos, others crumble under more sophisticated paraphrasing attacks. The study underscores a critical trade-off: embedding easily detectable watermarks often comes at the expense of text quality. Furthermore, the research shows that even subtle changes introduced by translating text and then back again can disrupt some watermarks. This raises concerns about the reliability of current watermarking techniques in real-world scenarios. WaterPark offers valuable insights into the strengths and weaknesses of different watermarking methods, paving the way for more robust solutions. The ongoing development of more resilient watermarks is crucial, ensuring we can trust the origins of AI-generated content and mitigate its potential misuse.