Claude Content Marking and What It Means for Provenance

A major model provider commits to marking its output

Anthropic has committed to the EU AI Act's Article 50(2) Code of Practice on Transparency of AI Generated Content. Starting August 2, 2026, new Claude models mark the content they generate by default. Existing models are being transitioned to the same standard. The marking applies across the Claude API, the Claude app, Claude Code, Claude Cowork, and Claude Tag, and it extends to Claude as offered through AWS, Google Cloud, and Microsoft Foundry.

This is the kind of commitment Limbo tracks closely. Provenance only works when the organizations producing content participate in marking it, and a frontier lab building marking into its default behavior is a meaningful signal for the field.

Two techniques, two purposes

Claude uses an imperceptible watermark for generated text. The watermark does not change how the text reads, and it is designed to persist through copy and paste and survive light editing. This is similar to the SynthID watermarking solution that Google launched recently for their text watermarking.

For image and vector files (svg, png, jpg), Claude attaches signed provenance metadata following the C2PA standard, the same standard Limbo builds on. That metadata lets a downstream tool confirm whether a file has been altered since it was generated.

Anthropic also says it is building tools that let users and third parties detect these watermarks independently, rather than asking people to take the marking on faith.

What the marks do not prove

Anthropic is clear about the limits, and those limits matter for anyone doing verification work. A detected mark shows that Claude touched the content at some point. It does not confirm sole authorship, and it does not rule out edits made after the fact.

The absence of a mark is even less conclusive. Older models predate the marking system. Heavy editing or paraphrasing can remove a text watermark. Short passages may not carry enough signal to detect. Converting a file to a different format can strip metadata entirely.

That last point is the one Limbo customers run into most. A screenshot or a re-saved image file can lose its provenance signal without losing its AI origin, which is exactly why relying on a single vendor's mark is not enough to establish trust.

Where this leaves the industry

Anthropic's move is an additional signal of real infrastructure behind a compliance requirement rather than a voluntary statement. Organizations building on Claude still need to assess their own obligations under Article 50 independently. And no single lab's marking system, on its own, gives a downstream platform full confidence in what it is looking at.

That is the gap Limbo exists to close: aggregating provenance signals across sources, models, and formats so platforms are not stuck trusting one label from one vendor. Anthropic's announcement is a good sign for the direction the industry is heading. It is also a reminder of how much verification work still needs a layer above any single provider's marks.

Want to see how Limbo verifies content provenance across sources? Get in touch.