Days after Anthropic revealed that text from its Claude AI would carry an invisible watermark, the firm released additional details on the technology. It rejected user ideas that machine-readable characters would be added to the output and confirmed no hidden elements would appear in the text. The system instead creates a detectable pattern in responses that requires a specific key to identify. Anthropic said the method is based on the SynthID-Text approach from a Google DeepMind paper published two years ago. The company emphasized that watermarking would not reduce content quality, creativity, readability, speed, or cost. Detection improves with longer passages because more text provides greater opportunity for the pattern to form. Code tends to show weaker watermarking than other text, while translations carry it since every word is selected by the AI. The watermark can indicate that Claude was likely involved but cannot confirm exact authorship or distinguish between original and edited content. Light edits may leave the mark intact, though replacing every word could remove it. The feature cannot trace output to any person, group, or session. Older models will gain watermarking in coming months. Anthropic said the step supports compliance with the EU AI Act, noting other major providers also joined the related transparency code in July.
