LLM Watermarking

LLM watermarking is a class of techniques that embed an imperceptible statistical signature in text generated by large language models, so that the output can be reliably identified as machine-produced by a downstream detector without altering meaning, quality, or readability. Modern LLM watermarking approaches typically bias the model's token-selection process during decoding — for example, by partitioning the vocabulary into "green" and "red" token groups using a cryptographic key and steering sampling toward green tokens — so that green-token patterns in the output betray the model's identity. The signature travels with the text when it is copied or lightly edited, and can be detected by anyone holding the corresponding key. LLM watermarking is distinct from older text-watermarking approaches for human-authored documents, and from cryptographic signing of the model's response (which proves identity at the message level but does not survive copying or paraphrasing). Regulatory pressure — notably the EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content — has made production-grade LLM watermarking a near-term deployment requirement for major model providers.