Anthropic Deploys Subtle Word-Choice Watermark in Claude AI Text
Anthropic said it has introduced a subtle word-choice watermark for text generated by its Claude models, subtly altering synonym selection to embed a machine-readable pattern while preserving readability, in a move aimed at complying with upcoming EU AI regulations. The company announced the change from its San Francisco, California office. Earlier, the firm announced that it would embed an invisible, machine-readable watermark in all Claude outputs worldwide to satisfy the EU AI Act, explaining that the system uses a keyed random selection of words (The Verge).
Anthropic initially planned to begin the rollout on August 2 (Fortune) and noted that the marks could also appear in lightly edited human-written text (Ars Technica). The introduction sparked criticism on social media and prompted some users to cancel their Claude Max subscriptions (TechCrunch; Business Insider).
