Claude’s text will soon carry a hidden signature. Anthropic says Claude models launched on or after August 2, 2026, include an “imperceptible watermark” in their output to comply with the European Union’s AI Act.
The EU’s rules require synthetic audio, image, video, and text to include machine-readable marks. Other major model developers signed the same Code of Practice and will implement their own watermarks, so Anthropic is not volunteering for a lonely regulatory experiment.
Anthropic announced the details on August 17, after publishing blog posts explaining the system from August 13 through August 15. The company also plans to add support for older Claude models and release a detection tool soon.
How Claude’s watermark works
Claude’s text marking system is “a version of the SynthID-Text approach” developed by Google DeepMind. Google Gemini has supported SynthID-Text since 2024, giving the method an existing place in the generative AI landscape.
The system does not attach a visible label or alter the appearance of a response. Anthropic states, “Watermarking does not impact the quality of Claude’s output,” adding that “A watermarked response is indistinguishable from an unwatermarked one to a reader.”
Instead, the watermark uses low-stakes choices made during text generation. Claude still makes those choices at random, but the source of randomness changes when watermarking is active, creating a detectable pattern across the response.
Those choices depend on a key and the words that came before them. Anthropic describes the result this way: “The watermark pattern is undetectable to the reader, but is detectable to anyone who has a key that encodes it.”
The watermark travels with copied-and-pasted text, which gives it a practical advantage over marks tied to a specific interface. A response can move from Claude into another document and retain the pattern without displaying anything unusual to the person reading it.
Editing can weaken the signal, but not erase every trace
Anthropic says light editing probably will not remove the watermark completely. A complete rewrite where every word is replaced can remove it, creating a clear limit for detection systems: preserved text is easier to inspect than text that has been rebuilt from scratch.
Code presents another limitation. Watermarking can be applied to code, but it will have less of a watermark than other text because code leaves fewer arbitrary choices during generation.
Anthropic explains that “in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code.” That makes comments a stronger target for marking than the code’s fixed syntax, where changing a word can break the result.
The planned detection tool will make the hidden pattern useful beyond Anthropic’s internal systems. Pangram also offers AI detection approaches, but the Claude system is built around a specific key and a watermark pattern generated during output.
OpenAI is subject to the EU AI Act requirements, and other major model developers have signed the same Code of Practice. That turns machine-readable marking into a shared industry requirement rather than a feature unique to Claude.
The practical effect is easy to understand: Claude output can look untouched while carrying information that a compatible detector can identify. Readers see ordinary text; systems with the key see a pattern. The watermark is invisible, but the compliance paperwork certainly is not.
Based on
- How Claude Watermarks AI-Generated Text — magazine.sebastianraschka.com
- 3 Arguments for and Against AI Watermarks – Business Insider — businessinsider.com
- Anthropic explains how Claude’s invisible text watermarks will work | The Verge — theverge.com
- Anthropic shares more details about how Claude’s new watermarks will work | TechCrunch — techcrunch.com
- What does it mean to put a ‘watermark’ on AI text? | CNN — cnn.com
