Image: The Decoder / the-decoder.com Anthropic Is Now Watermarking Everything Claude Generates
Starting with models released after August 2, 2026, Anthropic embeds invisible text watermarks and C2PA metadata into all Claude outputs — worldwide, not just in the EU.
Anthropic announced today that Claude models released on or after August 2, 2026 automatically watermark their outputs. The company is using two separate approaches: an imperceptible signal embedded in generated text, and C2PA provenance metadata attached to image files. Both apply globally, not just to users in the European Union.
The trigger is Article 50 of the EU AI Act, which took effect August 2 and requires AI companies to mark AI-generated or AI-edited content in machine-readable ways. Anthropic is complying — and extending the requirement worldwide rather than running a separate EU-only implementation.
What Gets Marked and How
Text watermarking works at the model level. The signal is woven into output as it’s generated, which means it travels with copy-pasted text. Anthropic says it “may persist through some editing” — exact durability under heavy rewriting is unspecified. The company has not said how much editing it takes to scrub the mark.
For files, Anthropic is using C2PA, an industry standard for recording a digital asset’s origin and edit history. The supported formats are SVG, PNG, and JPG. The metadata is digitally signed, so tampering can be detected — but only if the file stays in a C2PA-aware pipeline. Convert it, screenshot it, or open it in an editor that doesn’t support C2PA, and the metadata is gone.
Coverage includes every Claude interface: the Claude.ai platform, the API, Claude Code, Claude Cowork, and Claude Tag. Cloud partners — AWS Bedrock, Google Cloud Vertex, Microsoft Azure — may not support all marking types depending on their integration.
The Limitations Are Real
Anthropic is upfront about what watermarks cannot do. A detected watermark means the content may have been processed by Claude, not that Claude is the original author. Claude is widely used for editing, translation, and summarization — plenty of marked content will be human-written text that Claude touched.
The reverse is also true: missing a watermark doesn’t confirm something is human-written. Models released before August 2 don’t have marking support, and older models are getting it added gradually through a transition period. Short text may not have enough structure to reliably carry the signal.
Detection tools are coming, though Anthropic hasn’t given a timeline for when developers and third parties will be able to verify marks.
Not Just Anthropic
Google, Meta, Microsoft, OpenAI, and Synthesia have all signed onto the EU’s voluntary transparency code with similar commitments. The EU AI Act’s marking requirement is pushing the entire industry toward some form of output labeling, though the technical implementations vary by company.
Whether text watermarking actually works in practice is a separate question. The approaches described by Anthropic and others are statistical — they shift word distributions in ways that are detectable by machines but invisible to readers. Academic research has shown these signals can be degraded or removed with enough effort. Anthropic’s own guidance acknowledges the limitation: the watermark “may persist through some editing,” which leaves a lot of room for ambiguity.
For developers using the Claude API, nothing changes about how requests work. The watermarking happens inside the model. There are no new parameters to set and no opt-out.
Sources: