Anthropic is introducing a hidden watermarking system designed to help identify text generated by its AI model, Claude. Unlike traditional watermarks, this system will not add visible symbols, special characters or noticeable changes to the content.
Instead, Claude will create a subtle pattern through the way it selects words. Special detection systems will then be able to analyse that pattern and determine whether Claude was involved in producing the text.
Hidden Watermark Will Be Created Through Word Choices
According to Anthropic, when Claude has several equally suitable words to choose from, the system can adjust the underlying randomness used to make that decision.
Repeated across a piece of writing, these choices can create a distinctive statistical pattern. A detection system with the appropriate key can then identify the watermark.
Importantly, Claude will not be instructed to use unusual or awkward words simply to create the watermark. Anthropic says its internal testing found that the system does not significantly affect the quality, creativity or readability of generated content.
The technology is based on an approach related to Google DeepMind's SynthID-Text system, which was introduced in 2024.
The Watermark Will Not Prove Everything
Anthropic's watermark is intended to indicate that Claude likely played a role in producing a piece of text, but it will not provide absolute proof of its origin.
The system will not reveal the identity of the person who generated the content, their organisation or their conversation with Claude. It also will not tell users whether another AI system was involved.
Detection may be more difficult with very short pieces of text because there are fewer words available for the system to analyse.
The watermark may also be less effective for content such as factual writing, proofreading and computer code, where there may be fewer opportunities for the model to make meaningful word choices.
Can the Watermark Be Removed?
Anthropic says the watermark may survive certain minor edits, but completely rewriting the content can remove the detectable pattern.
Interestingly, translations produced by Claude can also retain the watermark because the model is still making word-selection decisions while producing the translated text.
This makes the technology different from conventional AI detectors. Traditional AI detection tools generally attempt to identify stylistic or statistical characteristics associated with AI-generated writing, whereas Anthropic's system is designed to identify a watermark specifically associated with Claude.
Anthropic is also developing a detection API that could allow users and third-party services to check text for the watermark.
Anthropic Is Also Using C2PA for Images
The company's efforts are not limited to text. Anthropic is also using C2PA content credentials for supported files created or modified with Claude.
These credentials can be included in formats such as PNG, JPG and SVG. The files can contain cryptographically signed metadata indicating that Claude was involved in creating or processing the content.
The metadata is designed to identify the involvement of the AI system without revealing personal information about the user.
C2PA is an open standard already being adopted by camera manufacturers, technology companies and image-editing software providers to establish the provenance of digital content.
Why Is Anthropic Introducing the System Worldwide?
Anthropic is rolling out the watermarking technology globally rather than restricting it to particular regions. One reason is that the company currently does not have a reliable way to limit the system to specific geographic locations.
The move also relates to growing regulatory requirements around transparency in AI-generated content.
Anthropic has signed the European Union's Code of Practice on Transparency of AI-Generated Content, joining numerous other organisations working on standards for identifying AI-created material.
The company also plans to add watermarking capabilities to models launched before August 2, 2026, over the coming months.
What Does This Mean for Claude Users?
The new system does not mean that every piece of Claude-generated text will automatically be identifiable with complete certainty. Detection depends on factors such as the length and type of the content and whether the original text has been substantially rewritten.
However, it represents a significant shift toward making AI-generated content more traceable without visibly changing the text.
In the future, users may increasingly be able to check whether content was generated or processed by a particular AI model without relying solely on conventional AI-writing detectors.
