FuriaHubAnthropic will watermark Claude text with a SynthID-Text variant
Future Claude models will embed an invisible statistical pattern in generated text to meet the EU AI Act. Anthropic says quality, cost, and identity of the user are unaffected.
On August 14 Anthropic published a technical note on how Claude will watermark generated text. The Verge and TechCrunch followed on August 15–17. The company said it is using “a version of the SynthID-Text approach,” the open method Google DeepMind described in 2024, to comply with the EU AI Act’s machine-readable marking rules that took effect August 2.
How the text mark works
Watermarking biases low-stakes next-token choices so a detector with the right key can score the likelihood that Claude wrote the passage. Anthropic says the pattern is not visible to readers, does not change quality or content in any practical way, does not add tokens or cost, and carries no identifying information about a person, organization, or chat.
Light human edits leave little for the watermark to attach to. Code is marked less than prose, because working programs leave fewer arbitrary word choices; comments can still carry a mark.
Scope and images
New Claude models will mark output from day one. Models launched before August 2 have a transition period; Anthropic is adding marks to those over the coming months. Marks apply globally because the company says it does not yet have a durable way to scope them by region. Generated files get C2PA provenance metadata where the format supports it. Anthropic also said it will ship a detection API.
- Surfaces named: Claude Platform (API), Claude, Claude Code, Claude Cowork, Claude Tag
- Text: SynthID-Text-style statistical watermark
- Files: C2PA where supported
Takeaways
- Driven by the EU AI Act and the July 2026 Code of Practice (~190 signatories)
- No opt-out described; applied globally at launch of new models
- Detection tools for third parties are promised, not yet documented in the note
Source: Anthropic
