30 September 2026 · 3 min
AI texts and traceability: Anthropic's invisible watermark
Invisible trace
Anthropic introduces a digital watermark in texts generated by Claude to comply with the European AI Act.
Exteryo Report editorial staff
How the invisible trace works, allowing the origin of contents to be verified without altering their quality.
The short answer
The traceability of contents generated by artificial intelligence becomes structural:
- PatternAnthropic inserts an imperceptible pattern in Claude's choice of words
- Verificationthe watermark allows analysis software to verify the synthetic origin of the text
- Transparencythe system responds to the new European provisions on transparency
The short answer
The trace holds up until the text is rewritten almost entirely.
Why it's not simple
The introduction of a mark in texts generated by artificial intelligence has sparked a heated debate among users.
There is more.
Why it's not simple
The choice follows the new provisions of the European AI Act, aimed at ensuring maximum transparency and traceability of automated production.
Chapter 1 of 4
How does the digital watermark work in Claude's texts?
Invisible pattern
Anthropic's watermark is inserted when the model selects words, creating a pattern invisible to human reading.
The source
The technology developed by Anthropic acts directly during the response composition phase. When the model selects the sequence of words, it weaves a recurring structure into the discourse.
Note
The system does not use the mechanisms that usually denote the use of AI, such as the use of em dashes or reinforcing constructs.
Invisible pattern
This pattern is imperceptible to human reading, but perfectly identifiable by dedicated analysis software.
The trace
This digital fingerprint allows the synthetic origin of the document to be verified without altering the fluency.
1 of 2
Readable text
2 of 2
Invisible pattern
Chapter 2 of 4
What are the limits of the watermark against removal attempts?
Limits
The invisible trace resists simple evasion attempts, but yields to an almost complete rewriting of the text.
Attempts
More experienced users look for methods to bypass the watermark, such as passing the text through another competing chatbot or rewriting it. Anthropic points out that other major developers will implement their own watermarks, making switching between chatbots ineffective.
Limits
Removing the watermark is only possible with an almost complete rewriting of the text generated by Claude.
Rewriting
- PassingPassing the text to other chatbots (ineffective if all adopt the watermark)
- PartialPartial rewriting (the pattern resists)
- CompleteAlmost complete rewriting (removes the watermark, but cancels the advantage of AI)
Limits
At that point, the company points out, it is worth asking whether we are still talking about a product generated by artificial intelligence.
Chapter 3 of 4
How can companies demonstrate compliance with the AI Act?
Compliance
The adoption of digital watermarks responds to the provisions of the European AI Act to ensure transparency and traceability.
Transparency
Anthropic's choice follows the new provisions of the European AI Act. The goal is to ensure maximum transparency and traceability of automated production.
Before moving on
When was this novelty announced by Anthropic?
Control
To facilitate control, Anthropic plans to release an API for watermark detection.
Chapter 4 of 4
What differences exist between tracking generated texts and computer code?
Code
In coding, the need to generate working code limits the applicability of the watermark compared to normal texts.
Priority
The watermark for those using Claude in coding will be less applicable than for normal texts. The priority is to generate working computer code, even before it is recognisable.
Code
However, a watermark could be applied in areas where there is an arbitrary choice between specific words or terms, such as in comments.
Code
In these cases, the watermark will have a negligible effect on the code actually produced.
In one sentence
Removing the watermark requires an almost complete rewriting: at that point, is it still an AI-generated text?
Exteryo Report
Takeaway
What to know about Anthropic's watermark.
0 of 3 checked
Write with us
Want to write for Exteryo Report?
If you would like to write or publish an article, or collaborate with us, email the editorial team or get in touch through the contact form.
Sources
Images and animations: illustrative objects generated with artificial intelligence. They do not depict real objects or documents mentioned in the article.
ExteryoExteryo Report · 30 September 2026