Claude Is Now Leaving Invisible Fingerprints In Its Text

This post details a method for embedding invisible statistical watermarks into text generated by large language models like Claude. The technique leverages the model's token generation process by biasing the selection of tokens based on a secret key and a pseudo-random number generator. This bias is applied such that the resulting text remains statistically indistinguishable from normally generated text to a human reader, but a detector can identify the watermark by analyzing the token probabilities. The paper cited demonstrates that this watermarking scheme is robust to various forms of text manipulation, including paraphrasing and summarization, making it a viable tool for identifying AI-generated content.