Solis: Claude adding Watermarks for AI generated content
Key Takeaways:
Anthropic announced that all text from Claude models that launched on August, 2nd 2026, will have an embedded watermark to be identified under EU AI Act
Claude will further provide documentation and tools in cooperation with partners to use for detecting their watermark
Still, they also speak of several limitations
How does it work?
A Claude model will subtly shift word choices during generation using a secret cryptographic key, creating a statistical pattern that is invisible to humans but detectable with the matching key
Foot build a streamlit app based on the paper of Kirchenbauer et al. 2023
Takeaway I: These marks are tough. Delete a third of the words at random and the text is still detectable.
Takeaway II: A clean result proves nothing about whether text is human.
What will be the impact of watermarks?
Therefore, the watermark itself will not be a reason for devaluation, the focus remains on quality produced by humans, AI or with the help of AI - still, the watermark might point for systems to re-assess the quality and effort
Haynes was surprised about the outrage on watermarking as Google's SynthID was introduced for images already back in 2023
Kunisch also stresses that there is nothing to do about it for SEO, GEO or AI Search experts - other than focussing on good quality content
Beus outlines that AI watermarked content is the new normal and users might be going to focus on human generated content
Sources:
