How Claude marks AI-generated content
Points and comments are a snapshot, not live.
Anthropic details how Claude embeds imperceptible watermarks in AI-generated text worldwide.
Anthropic's support article explains that Claude embeds an imperceptible watermark directly into generated text at the model level. The watermark does not change meaning, quality, or readability, and persists through copying and some editing. It applies to output from supported models wherever Claude is offered. Generated files will include digitally signed provenance metadata where supported. Content may not carry a detectable mark if the passage is very short.
What commenters are saying
Commenters focus on technical mechanisms, with several pointing to the "red token, green token" approach from an arXiv paper, where the model biases token selection based on a hash of the previous token. Two camps emerge: those who believe watermarking is easily circumvented (e.g., via grep or sanitizers) and those who argue that many users will not bother. Some note that Claude's distinctive writing style may already serve as a fingerprint. There is concern about false positives and the risk of people adopting LLM-isms.