- Anthropic introduces invisible watermarks for Claude AI text, which are woven directly into the text and do not alter its meaning, quality, or readability.
- The watermarking will automatically apply to all Claude models launched in the EU on or after August 2, regardless of the service provider used.
- Anthropic plans to extend watermarking to older models and roll it out globally.
- Detection tools for the watermarks are currently being developed by Anthropic and will be unveiled soon.
Anthropic's new invisible watermark technology for its Claude AI models aims to enhance text authenticity and traceability. The watermark is embedded directly into the text, remaining undetectable and preserving the original meaning and quality.13
This feature is effective for all Claude models launched in the EU from August 2, coinciding with the implementation of new regulations. The watermark survives copying, pasting, and light editing, making it difficult to bypass, although it may fail with very short texts or extensive edits.2
Anthropic plans to extend this watermarking to older models and is developing detection tools to identify these watermarks, which will be available soon. Notably, the watermark will be applied to all outputs from Claude, including translations and summaries, even if the AI was not the original author.4
The company is taking a cautious approach, implementing watermarking despite certain exceptions in the Code of Practice that allow for assistive functions in standard editing. This proactive measure underscores Anthropic's commitment to responsible AI deployment.
As the technology evolves, the implications for content authenticity and copyright protection will be significant, especially as the global rollout progresses.
“The watermark survives copy-paste and light editing but fails on very short text or heavy edits. Anthropic will watermark even assistive outputs like translation or proofreading, going beyond the EU Code's exception, and plans detection tools soon.”
