Anthropic watermarks every Claude model released after August 2, worldwide
Anthropic now skews Claude's word choice against a secret key so output is statistically detectable, a SynthID-Text-style mark on every model released after August 2 and applied worldwide because the EU AI Act feature cannot be limited by region. John Gruber argues the quality cost is real since no two synonyms carry identical meaning, and tools such as Declaude strip the mark by paraphrasing. The mark survives ordinary editing and travels into documents built on older templates, making the AI share of a contract verifiable in a fee negotiation.
Source: the-decoder.com ↗
When Claude picks between "overcast" and "grey" based on a secret watermark key rather than semantic precision, text quality inevitably suffers.
John Gruber, Daring Fireball
Why this matters
- → Detectable AI text becomes enforceable in contracts and fee disputes
- → Quality tradeoff: watermark forces suboptimal word choices for statistical marking
- → Global enforcement of EU rule creates friction for users worldwide
The watermark tax