Anthropic details plans for invisible Claude watermarking under EU AI Act rules
Following the announcement, earlier this month, that Claude-generated text will include invisible watermarks, Anthropic has provided further details about the topic.
The company has confirmed that the implementation of watermarks is happening to comply with the EU AI Act.
They also state that watermarking will not change output quality, content, or cost, and readers will not notice any difference.
Finally, no hidden characters or extra tokens are being added, and watermarks cannot be traced to specific users or chats.
Ro
- AI Chatbot
- Freemium
- Proprietary
Related news
External links
- We’ve written an FAQ to answer some of the questions we've received about watermarking. Anthropic on X • Official source
- How Claude's text watermarking worksAnthropic News • Official source
Gu
Comments
Credit where due — this addresses two concerns I raised under the earlier announcement: no hidden characters or zero-width tricks, and no tracing back to specific users or chats. That rules out the nightmare scenario of paste-detectable fingerprints following people around, and it sounds like statistical token patterning rather than steganography, which is the right way to do it.
But the open question just moves. Statistical watermarks need a minimum run of untouched model-chosen tokens to be detectable — short snippets and heavily edited text fall below the threshold. So the labeling will be strongest exactly where nobody disputes it (long, unedited generations) and weakest where the real arguments happen (mixed human+AI text). Curious whether EU AI Act auditors will accept that gradient, or start demanding certainty where the math doesn't provide it.