The article addresses a growing problem in the age of LLMs: machine-generated text can closely mimic human writing, which makes it harder to verify who or what produced a given piece of content and raises risks of misinformation, impersonation, and unauthorized reuse. A common defense is text watermarking, which hides a detectable signature in generated text. However, most existing methods embed the signature at the level of individual words or word-level statistics, and these signals tend to break down as soon as the text is paraphrased or lightly edited. The article argues this fragility is a serious practical weakness, since paraphrasing is easy and widely available, and it proposes a different place to hide the watermark: not in the words, but in the underlying meaning of each sentence. The article's main objective is to introduce and evaluate SWAN, a framework that embeds watermark signatures into the semantic structure of sentences using Abstract Meaning Representation (AMR), a graph-based way of representing the core meaning of a sentence independently of its exact wording. The central claim the article tests is that if a signature lives in the meaning structure rather than the surface text, then any paraphrase that preserves meaning should automatically preserve the watermark. The article evaluates whether this approach can match existing methods on ordinary watermarked text while being substantially more robust when that text is rewritten. Read with an AI tutor: chapterpal.com/s/ra5wruq4/sw…