Anthropic Plans Watermarks for AI-Generated Content from August 2026
Set Trending Topics as a preferred source on Google.
Anthropic plans to add machine-readable labels to content generated by its Claude models. The move is based on the company’s commitments under the Code of Practice on Transparency of AI-Generated Content for Article 50 of the EU AI Act. It applies to new Claude models launched in the European Union on or after August 2, 2026.
These models are expected to support marking from launch. According to Anthropic, the approach will apply worldwide and across different access points, including Claude applications, the API, and services offered through cloud partners such as AWS, Google Cloud, and Microsoft Foundry. This is conditional on the model and feature in question supporting the relevant marking method.
Invisible Watermarks for Text
For text, Anthropic plans to use invisible watermarks. They will be embedded directly in generated text and are intended not to alter its meaning, quality, or readability. Because the watermark is part of the text itself, it may remain present when text is copied and pasted elsewhere and can survive some editing.
Provenance Data for Images and Other Files
For supported file formats such as SVG, PNG, and JPG, Anthropic also plans to add digitally signed provenance metadata. This metadata follows the open C2PA standard. It is intended to indicate that a file was processed by Claude and to make tampering detectable. Its availability may, however, be limited depending on the platform and file feature.
Anthropic also says it is developing tools that will allow users and third parties to detect the watermarks and metadata. Technical details about the detection process have yet to be published.
Detection Is Not Proof of Full Origin
The marking does not provide conclusive proof that content was entirely generated by AI. A detected watermark indicates only that Claude may have processed the content. A person may, for example, have supplied the original ideas or text while using Claude only for proofreading, translation, summarization, or file conversion. Marked content can also be modified, shortened, or combined with other material afterward.
Conversely, the absence of a detected mark does not mean that content was not created or processed with AI. Watermarks may be absent from older models that do not yet support marking. Extensive editing, paraphrasing, translation, or very short passages can also make detection more difficult. For files, metadata can be removed through conversion, re-saving, screenshots, or similar actions. In addition, not every platform, feature, and file type will support every kind of marking.
Older Models to Follow Later
A transition period applies to Claude models released before August 2, 2026. Anthropic says it is working to add marking support to those models as well, but has not yet announced a specific timeframe.
