Anthropic: Claude Watermarks Use Google's SynthID-Text
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic revealed Claude's text watermarking uses a version of Google DeepMind's open-source SynthID-Text, which embeds detectable patterns by manipulating the probability distributions behind word choices
- The watermark mechanism steers low-stakes word picks — like choosing between "overcast" and "grey" to complete a sentence — by swapping an arbitrary random number generator for a key-based one, leaving a pattern readers cannot detect
- Anthropic says the watermarks will not raise Claude's price for users or affect the quality or content of its outputs
- The text marks, paired with new C2PA support for Claude-processed images, are being introduced to satisfy the EU AI Act's requirement that synthetic text, audio, images, and video carry machine-readable markers of AI origin
- Google's Gemini has supported SynthID-Text since 2024, making Claude the second major model to adopt the approach; OpenAI has not disclosed any text watermarking roadmap but will also be bound by the same EU rules
Why it matters: Anthropic and Google now share a technical standard — SynthID-Text — for meeting the EU AI Act's transparency mandate, leaving OpenAI as the conspicuous holdout with no announced watermarking plan and a compliance deadline it cannot ignore.
Ask SkimNews
