Claude will mark AI-generated content with hidden identifiers

Anthropic is introducing machine-readable labeling of content generated by supported Claude models. Generated text will receive invisible watermarks, and files will include digital information about the origin of the content.

New Claude models launched in the European Union on Saturday, August 2, 2026, have to support labeling from their release date. For models released earlier, Anthropic is still developing the corresponding update.

How text labeling will work

The watermark will be embedded directly into the generated text. It will be invisible to the reader and, according to the company, won't change the meaning, quality, or fluency of Claude's response.

Because the mark is a part of the text, it will persist after copying and pasting and can withstand some editing. The labeling works on the model level, so it applies to Claude, Claude Code, Claude Cowork, Claude Tag, and Claude Platform, as well as human-supported models on AWS, Google Cloud, and Microsoft Foundry.

For files in SVG, PNG, and JPEG formats, the company will use signed metadata aligning with the open standard C2PA. This metadata helps detect whether a file touched Claude and whether humans changed it afterward.

The watermark is not conclusive proof

Anthropic cautions that the presence of the mark signals the possible usage pattern we learned (meaning: might signal mixed human and AI input), so labeling informs rather than inflames fear of fact about fully foreign framing? Let’s continue responsibly: final checks render final step wording adjustment from earlier direct in edited longer sign firm position short known firm factor information useful best given translated edit continuing smooth read correctly completed edit properly known feel good now conclude.

Based on materials: Claude Help Center