Anthropic is implementing a global text watermarking system for all Claude models to comply with the EU’s Code of Practice on Transparency of AI-Generated Content. Rather than using invisible Unicode characters, the method utilizes semantic steganography by manipulating token selection during inference. Specifically, the system employs a secret key to assign words to "green" or "red" lists; the model is then probabilistically nudged to select tokens from the green list more frequently than the red list. This creates a statistical fingerprint detectable only by the provider holding the secret key. While Anthropic and Google (via SynthID) claim this technique does not significantly impact readability or quality—citing studies where human feedback differences were statistically insignificant (e.g., a 0.01% difference in thumbs-up rates)—critics argue this is a fundamental perversion of writing. Because the algorithm must occasionally bypass the most precise or semantically appropriate word to satisfy the "green list" requirement, it inherently prioritizes statistical signaling over linguistic perfection. This approach applies to all generated text exceeding 200 tokens, potentially flagging human-edited prose or proofread text as AI-generated if the model's contributions are substantial enough to trigger detection thresholds.