ทันเอไอ

Anthropic details 'invisible watermark' for Claude text to comply with EU AI regulations

Anthropic adopts Google DeepMind's SynthID-Text technology to embed statistical patterns into Claude's outputs, enabling AI content detection in compliance with the EU AI Act.

Anthropic, developer of the Claude chatbot, has detailed its 'invisible watermark' system embedded in all text generated by Claude to comply with the EU AI Act's transparency requirements (Article 50), which mandate that AI-generated content be identifiable.

Interestingly, Anthropic did not develop this technology in-house, instead using SynthID-Text—an open-source watermarking technology developed by direct AI rival Google DeepMind.

The system works not by hiding special characters or metadata within the content, but by adjusting token selection probabilities. When the model has multiple synonymous words to choose from, the system selects words based on a predefined statistical pattern. While indistinguishable to human readers, software with the secret key can trace whether the text originated from AI.

Anthropic confirms this watermark does not impact content, meaning, creativity, or writing quality, nor does it increase token counts or user costs. The watermark persists even if text is copy-pasted or slightly edited, though it has limitations: heavy rewriting, paraphrasing, or overly short text can cause detection to fail.

This system will roll out globally for new Claude models launched after August 2, 2026, covering web interfaces, APIs, and Claude Code. Image files such as PNG and JPG will instead use metadata embedding based on the C2PA open standard.

Although mandated by European law, Anthropic is applying this watermark globally rather than limiting it to European users.