Anthropic has launched a machine-readable content-marking initiative for materials generated by its Claude models. This initiative combines invisible text watermarks with digitally signed provenance metadata for supported files.
This move follows Anthropic’s commitment to the transparency guidelines outlined in Article 50(2) of the European Union AI Act.
Anthropic Adds Invisible Watermarks to Claude
According to the plan, Claude models that are released in the EU on or after August 2, 2026, will support content marking from the outset.
Anthropic has stated that these measures will apply globally, not just to users in Europe, across various Claude services, including the Claude Platform API, Claude, Claude Code, Claude Cowork, and Claude Tag.
Additionally, the company is working on enabling marking capabilities for models released before the August 2026 deadline, which will be covered under the EU law’s transition period.
Anthropic’s approach uses two distinct mechanisms: embedded watermarks for AI-generated text and signed provenance records for supported files.
For text, Claude embeds an invisible watermark directly into the generated output at the model level. Anthropic asserts that this marking does not alter the visible wording, meaning, readability, or quality of the response.
Because the watermark is embedded in the text itself, it can persist when content is copied and pasted to other platforms and may withstand some forms of editing.
This model-level design is crucial for enterprise security and content moderation teams. It ensures that the watermark is applied regardless of how Claude is accessed, whether through an end-user application, a developer API, or partner-hosted infrastructure.
Furthermore, the same text watermarking will be in effect when supported Claude models are accessed through services like AWS, Google Cloud, or Microsoft Foundry. However, support for file-based metadata may vary depending on the capabilities of individual cloud platforms.
For supported file types, including SVG, PNG, and JPG, Claude will attach signed provenance metadata on accordance with the Coalition for Content Provenance and Authenticity (C2PA) open standard.
This C2PA metadata can confirm that Claude processed a file and provide signals that indicate whether the file has been tampered with. If the signed metadata remains intact, downstream tools can determine whether the file has been altered after the provenance record was added.
This mechanism can help security teams, publishers, and investigators distinguish between content with a verifiable Claude-processing signal and files lacking reliable provenance.
Anthropic has warned that a detected mark does not definitively prove that Claude created all underlying content.
For instance, a document may still carry a Claude watermark after the model has proofread, translated, summarized, reformatted, or otherwise processed human-authored material. Additionally, the absence of a watermark does not confirm that a human created the content.
Detection may also fail when the text is heavily edited, paraphrased, translated, mixed with other writing, or is too short to provide a reliable signal. File metadata can also be removed through processes such as conversion, re-saving, taking screenshots, or unsupported workflows.
For developers creating products powered by Claude, Anthropic emphasized the need for them to independently assess their own transparency obligations under the EU AI Act. The company plans to provide further guidance detailing its marking and detection methods in more technical depth.
Stop new phishing & malware before they compromise your business. Integrate live intel from 15K SOCs around the world
