Claude output is watermarked now: what Article 50 enforcement means for your product

Since August 2, every new Claude model embeds a machine-readable mark in the text it generates, at the model level, and attaches signed C2PA provenance to the files it produces. This applies globally, not only to users in the EU. The timing is not a coincidence: August 2 is also the day the EU AI Office began enforcing the transparency obligations in Article 50 of the AI Act, with fines of up to EUR 15 million or 3% of global turnover, and with the Code of Practice serving as a presumption of conformity for anyone who signs on. If your product ships Claude output to users, two things changed at once: what the law expects of you, and what your output silently carries.

What Article 50 enforcement actually requires

Article 50 is the transparency article. For a product team it comes down to two obligations:

• Tell people they are talking to a machine. A chatbot or voice agent has to disclose that it is AI unless that is obvious from context.
• Mark generated content in a machine-readable way. Text, images, audio and video produced by an AI system need a detectable marking so downstream systems can identify synthetic content.

The Code of Practice is the escape hatch: follow it and you are presumed compliant.

What Claude's watermark does and does not do

According to Anthropic's own description, as reported by Euronews and The Decoder, the text mark travels with copy and paste and 'may survive part of editing'. The C2PA provenance on files is a signed manifest. Both are lost in predictable ways:

• The text mark degrades under heavy editing and is not, in Anthropic's words, proof of authorship. It indicates that Claude produced text; it cannot tell you who prompted it or whether a human rewrote half of it.
• The C2PA metadata is stripped by format conversion, by screenshots, and by any editor that does not support the standard. A PNG exported from a tool that ignores C2PA is a clean file.

So the watermark satisfies the machine-readable marking half of Article 50 for content that goes straight from Claude to a user. It does nothing for the disclosure half, and it becomes unreliable the moment content passes through your own pipeline.

A model-level watermark covers the vendor's obligation. Your obligation starts where the content leaves the API, and that is exactly where the mark is weakest.

Compliance checklist for a product built on Claude

• Add the disclosure. A persistent, visible 'this is an AI assistant' line in every conversational surface. Do not rely on the watermark for this; it is invisible to the user by design.
• Decide what you do with the mark downstream. If your pipeline rewrites, summarizes or translates Claude output before showing it, assume the mark is gone and add your own marking or metadata at the output stage. Guardrails and output validation is where that step already lives.
• Preserve C2PA where files pass through. Audit every export path: image resizing, PDF generation, document conversion. Each one is a place provenance can be dropped.
• Know what your deliverables carry. White-label documentation, generated code comments, marketing copy: all of it now carries a detectable Claude mark if it was pasted unedited. That is a disclosure question for your clients and a record-keeping question for you; prompt versioning for a small team is how you keep the history of what was generated when.
• Log the agent surfaces. Anthropic extended its Compliance API to Cowork and Claude Code the same week, in beta for Enterprise. If you are on that plan, wire it in; it is your evidence trail for the coding-agent surfaces.

The honest limitation

Anthropic has not published how the text mark is detected, how robust 'some editing' is in practice, or who gets access to a detector. Without a detector you cannot verify your own pipeline, which means you cannot prove to an auditor that the mark survives your export path. The watermark is a good-faith answer to Article 50's marking requirement from one vendor, not a compliance program. Build the disclosure, the downstream marking and the logs yourself, and treat the model-level mark as a bonus you cannot depend on.