AI-generated content may soon come with a digital trail you cannot see but machines can read.
Anthropic says its new Claude models will support machine-readable marking at launch, as part of its commitments under the EU AI Act's transparency rules for AI-generated content. Here's how it works.
What is Anthropic changing? #
Claude models launched in the EU on or after August 2, 2026, will support machine-readable marks on generated content. Anthropic says the system uses two complementary methods: an invisible watermark embedded in text and signed provenance metadata attached to supported files.
How does the text watermark work? #
The watermark is embedded directly into generated text without changing how it looks or reads. The mark is designed to remain with the text when it is copied, allowing detection tools to identify that the content has passed through Claude.
What about images and other files? #
For supported formats such as PNG, JPG and SVG, Anthropic will use signed provenance metadata based on the C2PA (Coalition for Content Provenance and Authenticity) open standard. The metadata is attached to the file rather than embedded visibly in its content.
The Age of Anger #
31 Jul 2026 - Vol 05 | Issue 31
A raging generation makes the government relent. What's next?
Will the marking work everywhere Claude is used? #
Anthropic says the marks will apply across supported Claude models on Claude, its API, Claude Code, Claude Cowork and Claude Tag, as well as wherever Claude is offered globally. However, not every platform supports every type of marking. For example, embedded watermarks will be supported when Claude is accessed through AWS, Google Cloud and Microsoft Foundry, while signed provenance metadata may depend on the capabilities of each platform.
What about older Claude models? #
There is a transition period for models launched before August 2, 2026. Anthropic says it is working to add marking support to those earlier models as well.
Can these marks prove that Claude wrote something? #
No. And this is an important distinction. Anthropic says detecting a Claude mark shows that the content was processed by the system, but does not conclusively establish authorship or original provenance. Someone could use Claude to proofread, translate, summarise or convert material originally created by someone else. The resulting content could still carry a Claude mark.
Can a piece of AI-assisted content appear unmarked? #
Yes. Anthropic says the absence of a detectable mark does not prove that AI was not involved. Short passages, extensive human editing or stripping metadata while resaving a file can remove detectable signals.
So, what is the point? #
The goal is transparency, not an infallible AI detector. The marks give users and third parties another way to establish that content has passed through Claude, while acknowledging that digital provenance can be altered or lost. Anthropic says developers deploying Claude independently will still need to assess their own legal and compliance obligations. In other words: Claude won't necessarily tell you who wrote something. It will increasingly be able to tell you whether its AI has touched it.
(With inputs from ANI)