Anthropic explains how Claude’s invisible text watermarks will work
The Verge Jess Weatherbed ● Covered by 2 sources
Anthropic says Claude text will get invisible marks. It’s for Europe’s AI rules, and the marks are based on Google DeepMind’s open-source system.
Based on reporting by The Verge, Jess Weatherbed — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic has spelled out how it plans to watermark text from Claude without making the marks visible to readers. The company says the system it will use is a version of SynthID-Text, an open-source approach from Google DeepMind that hides detectable patterns in the probabilities behind word choices.
That matters because Anthropic is trying to line up with Europe’s AI transparency rules. The European Union’s AI Act requires synthetic audio, image, video and text to carry machine-readable marks, so this is less a product flourish than a compliance move.
Anthropic also says Claude-processed images will get support for C2PA, the content provenance standard used to track where media came from and what happened to it. Put together, the two features are meant to make Claude’s outputs easier to identify as machine-made or machine-handled.
The telling part is the choice of a watermark that stays invisible to people but remains detectable by software. That keeps the text readable and the policy people happier, which is the sort of compromise AI companies keep making as regulation gets real.
My take — AI-written commentary, not fact-checked reporting
This is the sort of move that sounds boring until you remember how much AI content is already drifting around looking human enough. Europe keeps forcing the industry to add receipts, and that’s healthy. The open question isn’t whether watermarking is perfect; it’s whether companies will do the minimum and call it governance.
Read more about this at: The Verge
