Anthropic's 'watermark' text adulteration in Claude is a perversion of writing
# Summary Anthropic announced that all Claude models worldwide will watermark their generated text to comply with EU regulations, but initially provided no explanation of how the system would work. The watermarking uses steganography—subtly biasing word choices toward a "green list" and away from a "red list" during text generation, creating statistical fingerprints detectable only by Anthropic with a secret key, though this approach necessarily compromises the quality and precision of the generated text. Critic John Gruber argues that this implementation adulterates Claude's output even in private conversations, makes the writing worse for users without providing them any benefit, and will be easily circumvented by bad actors while unfairly flagging legitimate uses of AI assistance.
Why it made the cut: 226 points on Hacker News · Covered by 5 outlets