Text Watermarking

Text watermarking is the practice of embedding an imperceptible signal into natural-language text so that the text can later be identified, attributed, or authenticated by a detector. The signal may be statistical (subtle shifts in word or character frequencies), syntactic (unusual but natural-looking phrasing patterns), or semantic (paraphrase-preserving transformations tied to a key). Robust text watermarks survive copying, light editing, and partial paraphrasing; fragile watermarks detect tampering but not copying. Modern LLM-era text watermarking focuses on machine-produced text and overlaps heavily with LLM watermarking, but the broader field also covers provenance marking for human-authored documents, plagiarism detection, and steganographic communication. Detection of an unmarked-to-watermarked conversion is asymmetric: a watermarked text always carries the signal, but absence of a detected signal does not prove the text is unmarked — only that no known watermark is present. Text watermarking is increasingly entangled with regulatory disclosure regimes that mandate machine-readable marking of AI-generated content.