Anthropic Fully Implements 'Invisible Watermarking'! Traceable Origins for Copied Text and Generated Images

Artificial intelligence (AI) giant Anthropic recently announced the full rollout of a 'machine-readable' content tagging mechanism across its Claude series models. From now on, all text content generated by Claude will be embedded with invisible watermarks imperceptible to the human eye, while image files will be attached with digitally signed provenance information.

According to Anthropic's official statement released on Tuesday (the 11th), this mechanism differs from most existing tagging methods that rely solely on file 'metadata'.

Claude's text watermarking is directly woven into the content itself. This means that even if users copy AI-generated text to other platforms or make partial edits, the watermark may still persist, significantly enhancing the ability to trace AI-generated content compared to previous solutions.

For image files, Anthropic employs a different approach: when Claude produces files in formats such as SVG, PNG, and JPG, the system automatically attaches digitally signed provenance information, adhering to the Content Authenticity and Provenance Alliance (C2PA) open standard widely adopted across the industry.

Any file bearing this signed label can serve as evidence that the content was processed by Claude, and it also helps determine whether the file has been tampered with afterward.

The key background prompting Anthropic to implement this measure is the company's signing of the 'AI-Generated Content Transparency Code of Practice' under Article 50(2) of the European Union's Artificial Intelligence Act.

Under the regulation, any new version of the Claude model launched within the EU from August 2, 2026, must have content tagging capabilities from its launch date.

Notably, Anthropic emphasizes that this tagging mechanism is not a region-limited measure solely for EU regulatory compliance, but will be universally applied across all Claude product lines and partner cloud platforms, covering global markets.

Services confirmed to be included are Claude Platform (API), the Claude application itself, Claude Code, Claude Cowork, and Claude Tag, as well as users accessing Claude through cloud partners such as AWS, Google Cloud, and Microsoft Foundry.

For existing models released before this regulation takes effect, Anthropic states it is currently working on retrofitting these 'legacy' models with tagging functionality, as the EU regulation provides a transitional grace period. Details will be announced progressively.

Alongside the rollout of the new mechanism, Anthropic has unusually detailed the system's limitations in official documentation, cautioning against treating it as an absolutely reliable identification tool.

The company notes that even if a Claude tag is detected, it only indicates that the content 'may have' been processed by Claude, without enabling 100% reconstruction of the full source.

For example, many users submit existing documents to Claude for proofreading, translation, summarization, or format conversion. While such outputs carry Claude's tag, the original ideas, arguments, or even data behind the text originate elsewhere. The content may also undergo further deletion, editing, excerpting, or merging with other materials after Claude's processing.

Conversely, 'no tag detected' cannot serve as proof that content is not AI-generated.

Anthropic lists multiple scenarios that could lead to missing tags, including: content generated by older models before the tagging feature was launched; text heavily rewritten, translated, or mixed with other content; passages too short to carry recognizable signals; files losing original metadata during format conversion, saving under a new name, or screenshot capture; or content originating from platforms or file formats not yet supporting this tagging technology.

For enterprises and developers integrating Claude technology into their own products, Anthropic specifically reminds them in its documentation that each party must still independently assess the specific transparency obligations imposed by Article 50 of the Artificial Intelligence Act on their products and services.

In other words, even though Anthropic provides the tagging tool at the technical level, downstream developers' compliance responsibilities are not automatically transferred or waived, and they must proactively clarify the scope of obligations they bear under the regulatory framework.