Anthropic is adding invisible watermarking to Claude-generated text and file output worldwide, not just in the European Union. The move comes as new EU AI Act transparency requirements take effect, but Anthropic says the marking will apply everywhere Claude is available.
The new system covers Claude’s consumer app as well as its developer and enterprise services, including the Claude API, Claude Code, Claude Cowork, Claude Tag, and Claude deployments through AWS, Google Cloud, and Microsoft Foundry.
Claude Will Get Invisible Watermarks
Anthropic says Claude models released from August 2, 2026 will support output marking at launch, with the company also working to add the capability to older models.
For generated text, Anthropic will embed an imperceptible pattern directly into the response. The company says the marking is designed to remain detectable after text is copied and pasted without changing its meaning, readability, or quality.
The system works differently for generated files. Supported formats such as PNG, JPG, and SVG can include signed provenance metadata indicating that the file was processed by Claude.
Anthropic Is Applying the Marking Globally
The watermarking will not be restricted to users in Europe. Claude output will receive the marking regardless of where the user is located, meaning users in the US, Asia, and other regions will receive the same provenance marking even though the EU’s AI transparency rules are the primary reason for the change.
Rather than maintaining separate systems for EU users and everyone else, Anthropic is applying the same output-marking system across Claude. This gives the company’s AI-generated content a consistent provenance layer across its global services.
How Claude’s Watermarking Works
Anthropic is using two different approaches depending on the type of output.
- For text, the watermark is embedded directly into the generated content as an imperceptible pattern that can be detected by software. It is designed to remain associated with the text even when the response is copied elsewhere.
- For files, Claude can attach signed provenance metadata. This works more like a digital record attached to the file, allowing compatible tools to determine whether the file was processed by Claude and whether its provenance information remains intact.
The marking is not necessarily something users will see by opening a file’s normal properties. Detecting it may require a compatible provenance tool or Anthropic’s own detection technology.
Claude Watermarks Can Be Removed
Anthropic’s system is not designed to make Claude output permanently traceable.
File metadata can be lost when a file is converted to another format, resaved, screenshotted, or otherwise processed. Similarly, text watermarking is not a guarantee that every piece of Claude-assisted writing will always be identified.
A watermark can indicate that content passed through Claude, but it cannot necessarily prove that Claude originally created all of the underlying material.
For example, someone could give Claude an existing article and ask it to summarize or translate the text. A resulting response could carry Claude’s marking even though the source material was created elsewhere. The opposite can also happen. Significant editing, translation, short passages, or older Claude models without marking support could make detection more difficult.
Why Anthropic Is Adding Watermarks Now
Article 50 of the EU AI Act requires providers of certain AI systems to ensure that synthetic content can be identified as artificially generated or manipulated. The transparency obligations took effect on August 2, 2026.
The EU’s approach favors multiple technical methods, including metadata, watermarking, fingerprinting, and other mechanisms, rather than relying on a single system.
Companies that fail to comply with applicable AI Act obligations can face significant penalties, including fines of up to €15 million or 3% of total worldwide annual turnover, depending on the violation.
Anthropic Joins the AI Watermarking Push
Anthropic’s decision puts Claude alongside other efforts to make AI-generated content easier to identify.
Google DeepMind has developed SynthID, which can mark AI-generated text, images, audio, and other media. OpenAI has also explored watermarking and provenance technologies for AI-generated content, although approaches and deployment timelines vary between companies.
Anthropic’s approach is notable because it combines text marking with provenance information for generated files and applies both systems globally.
For Claude users, the change should mostly happen in the background. The output will continue to look and behave normally, while compatible detection systems can use the added marking to determine whether content passed through Claude.
