Identifying AI-generated content is entering a new phase. Anthropic has introduced a system for Claude in which text produced by its assistant models carries a digital watermark that is invisible to an ordinary reader but designed to be recognized by machine systems. AI-written text no longer needs a visible label to be traceable.
The approach differs substantially from a conventional watermark. Unlike a logo on an image or a visible mark on a document, the Claude watermark is embedded within the text itself. According to Anthropic, it does not affect the meaning, quality or readability of the writing. The reader sees ordinary text, but it may contain a signal that helps identify its connection to an AI system.
A mark that travels with the text
A notable feature is that the watermark is not meant to stay confined to the original Claude conversation. Because the mark is part of the text, it can remain in place after the writing is copied into another document, a website or a social media post. Anthropic says the mark may also survive certain kinds of editing.
In practice, that could mean simply copying text from Claude and posting it elsewhere is no longer enough to remove every technical trace of its AI origin. Anthropic has not described the system as unbreakable, however, or claimed that it will necessarily survive every kind of modification.
According to Anthropic’s support documentation, assistant models released from August 2, 2026 are planned to include the marking capability from the start, while the company is working on adding it gradually to existing models. It would therefore be wrong to assume that every piece of text currently generated by Claude carries the same hidden watermark.
A different method for images
Images and other supporting files are handled differently from text. Files such as SVG, PNG and JPG can carry signed provenance metadata. The metadata follows the C2PA standard used across the industry and is intended to provide verifiable information about a file’s origin and the processing it has undergone.
That digital record can indicate that a file was processed by Claude and can help detect certain later changes to the file. But there is a fundamental distinction: the text watermark is embedded in the writing itself, while the C2PA information is attached to the file’s metadata. Taking a screenshot of an image, converting it to another format or resaving it in a way that discards metadata can therefore strip the provenance information. Treating file metadata and in-text watermarks as the same technology would be a mistake.
Regulatory backdrop
The change is tied to growing transparency demands in the AI industry. Article 50 of the European Union’s AI Act includes obligations to make certain AI-generated or AI-modified content identifiable in a machine-readable form. The aim of those requirements is that, in the future, distinguishing AI-generated material from human work will not rely only on appearance or guesswork.
