Every Claude model released since 2 August embeds a statistical watermark in its word choice to satisfy the EU AI Act. Lacking a regional switch, Anthropic applies the method worldwide.
The method builds on Google's SynthID-Text approach: during generation, the randomness source behind word selection is altered so that a statistically detectable pattern emerges. No visible or hidden special characters are inserted. Older Claude models are to be retrofitted over the coming months.
Anthropic says content, creativity and readability are unaffected. Writer John Gruber disagrees: if a model chooses between two synonyms by watermark key rather than by meaning, text quality must suffer. Anthropic concedes the marking is sparser in fact-heavy passages, where fewer alternatives exist. Paraphrasing tools can strip the marking.
What matters in practice is that the marking travels with the text. A contract may contain individual marked clauses while the rest is human-written; later contracts based on old templates carry the marking forward. Trade publication Artificial Lawyer considers this mostly unproblematic for law firms – the difficulty arises where clients have contractually barred AI use or where fees track effort.
What this means for decision-makers
- Establish which of your contracts and tenders exclude AI use – that exclusion is now technically verifiable.
- Define which classes of document may be produced with language models and which may not, and record the decision in writing.
- In advertising and agency work, expect clients to be able to evidence the AI share and to raise it in fee negotiations.
This story was produced automatically from the source named above and checked by software before publication. The image is symbolic and shows neither the event nor a real person. How this paper is made