Anthropic Plans Watermarks for AI-Generated Content from August 2026
Trending Topics Jakob Steinschaden ● Covered by 5 sources
Anthropic will mark Claude-made text and files starting with new EU models in August 2026. The tag survives some editing, so the source of a passage may stay visible even after copying.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic is preparing machine-readable labels for content made by Claude, with the plan tied to its commitments under the EU AI Act’s transparency rules for AI-generated content. The change applies to new Claude models launched in the European Union on or after August 2, 2026, and Anthropic says those models should support marking from launch.
The company says the system will not be limited to one product surface. It is meant to work across Claude apps, the API, and cloud partners including AWS, Google Cloud, and Microsoft Foundry, though only where the model and the feature support the relevant marking method.
For text, Anthropic is planning invisible watermarks. They are embedded in the generated text itself and are supposed to leave meaning, quality, and readability unchanged. Because the mark sits inside the text, it may travel with copied passages and can survive some editing.
Files get a different treatment. For supported formats such as SVG, PNG, and JPG, Anthropic plans to attach digitally signed provenance metadata using the open C2PA standard. The point is to show that Claude processed the file and to make tampering easier to spot, although the company says availability may depend on the platform and the file feature.
Anthropic is also building tools for users and third parties to detect both the watermarks and the metadata, but it has not published the technical details yet. And the company is careful about the limits: a detected mark does not prove something was fully made by AI, while no detected mark does not prove the opposite. Older models, heavy editing, paraphrasing, translation, short snippets, and file conversion can all complicate detection.
There is also a transition period for Claude models released before August 2, 2026. Anthropic says it wants to add marking support there too, but it has not given a timeline.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of boring: a concrete marking plan beats another grand speech about “trust.” But the real test is whether detection tools work outside Anthropic’s own nicely boxed world, because watermarks that vanish the moment someone screenshots a file are just compliance with extra steps. EU rules are pushing the industry toward provenance, and honestly, that’s one of the few bits of AI regulation that sounds like it was written by adults.
Read more about this at: Trending Topics