Anthropic is introducing machine-readable markings for content generated by Claude, including invisible watermarks in text and digitally signed metadata in supported files.
The change follows Anthropic’s decision to sign the European Union’s Code of Practice on Transparency of AI-Generated Content, which provides a framework for complying with AI Act transparency requirements.
Invisible Watermarks in Claude Text
For supported Claude models, generated text will contain an invisible watermark embedded directly into the output.
Anthropic says the watermark does not affect the meaning, quality, or readability of the response. Because it is embedded in the text itself, it can remain when users copy and paste the writing into another application and may survive some editing.
Supported files created by Claude, including PNG, JPG, and SVG files, will instead carry signed provenance metadata based on the C2PA industry standard. This can indicate that Claude processed the file and whether it has subsequently been altered.
The marking system applies worldwide wherever supported Claude models are available, including Claude, Claude Code, Claude Cowork, Claude Tag and the Claude API. Supported models accessed through AWS, Google Cloud and Microsoft Foundry will also receive text watermarks.
Users cannot opt out of the new markings, according to Forbes.
Existing Models Are Still Being Updated
Claude models launched in the EU on or after August 2, 2026 support machine-readable marking from launch.
Anthropic is still adding the technology to models released before that date, meaning not every existing Claude output will necessarily contain a detectable watermark yet.
The company also plans to release tools and technical documentation explaining how users and third parties can detect its watermarks.
Watermarks Can Still Disappear
Anthropic acknowledges that the system is not foolproof.
Heavy editing, paraphrasing, translation, or combining Claude-generated text with other writing could make a watermark undetectable. Very short passages may also be too small to provide a reliable signal.
File metadata can disappear when users convert or re-save files, take screenshots, or use software that strips metadata. Anthropic also warns that detecting a watermark only shows that Claude may have processed the content. It does not prove Claude originally wrote it.
That distinction has contributed to criticism from users who rely on Claude for proofreading or editing their own work. Others have raised concerns about how AI markings could affect coding workflows and attribution.
The EU says about 190 organizations had signed its transparency code by the end of July, including Anthropic, Google, Meta, Microsoft and OpenAI.
The post Claude Will Now Secretly Watermark All Text and Images appeared first on ProPakistani.