The company will use Google DeepMind’s SynthID-Text and release a detection API for EU AI Act compliance.
Anthropic published a blog post on Friday explaining how it will watermark text generated by its Claude chatbot. The company will use Google DeepMind’s SynthID-Text approach and plans to release a detection API to let organizations verify watermarked content.
The move follows Anthropic’s earlier announcement that it will watermark Claude output to comply with the EU AI Act’s Transparency Code. Many users have reacted negatively, with some on Reddit calling it a conspiracy and others on X claiming to cancel their subscriptions.
For builders, the key implication is that light edits will not fully remove a watermark, while a complete rewrite will. In code, watermarking will appear only in arbitrary choices like comments, because the model must produce functional code.
Anthropic plans to release the watermark detection API, but it has not announced a timeline. The company also clarified that watermarking differs from pattern-based AI detection tools that look for writing tells, such as those offered by Pangram.
What matters
- Anthropic explains how Claude’s watermarking works using SynthID-Text from Google DeepMind.
- Light edits to watermarked text will keep the watermark, but full rewrites will remove it.
- Watch for the launch of Anthropic’s watermark detection API for enterprise users.
Why it matters
Watch for the launch of Anthropic’s watermark detection API for enterprise users.
This GenAI News article was prepared in original wording using reporting and materials published by TechCrunch AI. Source reference: https://techcrunch.com/2026/08/15/anthropic-shares-more-details-about-how-claudes-new-watermarks-will-work/.
Drafted by the GenAI News review pipeline.