AI BriefingAnthropic adds invisible AI text watermarks to latest Claude models for EU AI Act compliance
·2026.10.22 09:00
Key point
The watermark uses SynthID-Text steganography to hide signals in word choices without adding tokens or revealing user identity.
Details
- Anthropic has implemented invisible AI watermarks in text outputs from its latest Claude models, including Claude Mythos 5.1, Claude Fable 5.1, Claude Opus 5.5, and Claude Sonnet 5.5, to comply with the EU AI Act.
- The technology relies on SynthID-Text, an open-source steganography project by Google DeepMind, which alters the randomness of token selection rather than adding hidden characters or consuming extra tokens.
- Google tested the watermark on roughly 20 million Gemini responses and found a 0.01% difference in thumbs-up rates, indicating the text quality remains imperceptibly unchanged to human readers.
- OpenAI has also announced support for watermarking via its "textGrain" technology, achieving 95% detection accuracy for passages containing 400 tokens, while Google has released a public website to verify non-text media watermarks.
- The EU AI Act requires AI providers serving the European market to mark AI-generated content as of August 2, 2026, a requirement also mirrored by California's SB-942 AI Transparency Act.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
OpenAI GPT-Rosalind with 1,946 Agents and Claude Scorpion Toxin Binder Design: Two Cases in AI Scientific Research
OpenAI nears $30 billion funding round at $1.25 trillion valuation led by UAE investors