Claude to get invisible watermarks for AI-generated content

Mumbai: Claude is getting a digital signature of sorts. Anthropic said it will mark content generated by its AI models to meet commitments under Article 50(2) of the EU AI Act’s Code of Practice on Transparency of AI-Generated Content.

The company has signed the code as a provider of both generative AI models and systems. Under its new approach, Claude models launched in the European Union on or after August 2, 2026 will support machine-readable marking from day one.

The system will use two methods: imperceptible watermarks embedded in generated text and digitally signed provenance metadata attached to supported files.

Anthropic said the markings will apply globally, not just to users in the EU. They will cover output from supported Claude models across Claude Platform’s API, Claude, Claude Code, Claude Cowork and Claude Tag, although some platforms and features may not support every type of marking.

For text, Claude will embed an imperceptible watermark directly into its output. The watermark is designed not to affect the meaning, quality or readability of the content.

Because the mark is embedded in the text, it can travel with content when it is copied and pasted elsewhere and may survive some forms of editing. Anthropic said the watermarking will operate at the model level, meaning it will apply regardless of which Claude product or interface generates the text.

The company is also working on tools that will allow users and third parties to detect these marks. Further details on how detection will work are expected in technical documentation.

For supported files, Claude will use a different approach. Generated formats such as SVG, PNG and JPG can carry signed provenance metadata based on the Coalition for Content Provenance and Authenticity (C2PA) open standard.

The metadata records information about a file’s provenance. Where the signed label is present, it can indicate that the file was processed by Claude and help determine whether the file has subsequently been tampered with.

The markings will also work when supported Claude models are accessed through cloud platforms such as AWS, Google Cloud and Microsoft Foundry. However, Anthropic said signed provenance metadata may not be available on every platform, depending on the features offered.

The new marking system applies to Claude models launched on or after August 2, 2026. Anthropic said it is also working to introduce marking support for models released before that date during the transition period allowed under the EU AI Act.

The company has not suggested that marking will provide an infallible way to identify AI-generated material. Instead, it describes the technology as an additional signal about how content was produced or processed.

That distinction matters because Claude can be used to edit, translate, summarise or reformat material created by humans. In such cases, the resulting content could carry a Claude mark even though its underlying ideas, text or data came from elsewhere.

Similarly, marked material may be edited, excerpted or combined with other content after Claude processes it.

Anthropic is also cautioning users against treating the absence of a watermark as proof that content was not generated or processed by AI.

A detectable mark could disappear after heavy editing, paraphrasing or translation. Very short passages may also not contain enough text for reliable detection. For files, metadata can be removed through format conversion, re-saving or even screenshots.

Content produced using older models, unsupported features or file types may also lack a particular marking mechanism.

In short, a mark can provide a useful clue, but no mark should not be treated as a clean bill of human authorship.

For organisations building products with Claude, Anthropic said developers should independently assess what Article 50 requires of their own products and services.

The company said it intends to support developers in meeting their transparency obligations by publishing more information about its marking and detection systems.

As AI-generated content becomes harder to distinguish from human-created work, Anthropic’s approach puts provenance at the centre of its compliance strategy. The bigger test now will be whether these invisible signals remain useful as AI content moves through increasingly complex editing and publishing workflows.