Large language models (LLMs) possess incredible abilities, but they've sparked concerns about copyright infringement due to their potential to generate copyrighted material. Researchers are exploring watermarking as a solution—embedding hidden signals within the LLM's output that are imperceptible to humans but detectable by algorithms. These watermarks, like the “UMD” and “Unigram-Watermark” methods, work by subtly altering the selection of words, making it significantly less likely for the model to reproduce copyrighted text verbatim. Tests show that watermarking can make generating copyrighted content tens of orders of magnitude less probable, while only slightly affecting text quality. The strength of the watermark can be balanced to find an optimal point between protection and generation quality. But, there's a catch! This very same watermarking makes it more difficult to detect copyrighted material within the training data itself. Methods for detecting whether a given text was part of the training set, known as Membership Inference Attacks (MIAs), see their success rates drop when watermarks are present. This presents a challenge: watermarking aids in preventing future copyright infringement during text generation, yet makes detecting past copyright misuse in the training data harder. The research suggests this challenge might be overcome. An “adaptive” MIA that understands the watermarking process could be developed, and preliminary studies show this approach could significantly enhance detection rates, even in the presence of watermarks. This research highlights the trade-offs and implications of watermarking LLMs for copyright law. While it's a promising tool against future infringement, new methods must be found to ensure continued auditing and address potential past training data misuse to improve overall copyright protection in the age of AI.