Anthropic to Add Invisible Watermarks to Claude's Text Outputs

HarperSCarroll · x · 2026-08-11

Anthropic is introducing watermarks for its Claude model's text outputs. Similar to Google's SynthID, this mechanism works by slightly shifting the probability distribution of the generated text, making it completely invisible to human readers.

Currently, the detection mechanism for this text watermark exists only within Anthropic, though the company states it is working to enable detection by users and third parties. For image outputs, Claude will add a signature to the metadata, which is notably easy to remove.

Related event: Anthropic Introduces Invisible Watermarks for Claude to Comply with EU Regulations(77 posts)→

Original post →

More from Models

Models channel →