Text generated by Anthropic’s Claude will now include invisible watermarks
Artificial intelligence company Anthropic will implement watermarks on all text created by its Claude model. The new feature, detailed on a support page updated on August 11, 2026, although motivated by European Union legislation, will be extended to all regions where the AI tool is accessible.
Anthropic reported that “Claude models released on or after August 2, 2026 support watermarking from launch.” Additionally, the company is in the process of integrating this functionality into older Claude models that were launched before that date.
These textual watermarks will be inserted in all Claude outputs, regardless of the platform used, including Claude (chat), Claude Platform (API), Claude Code, Claude Cowork and Claude Tag. The company reinforces that the “embedded watermarks will be applied to all generated text”.
More on this story: US Senator launches package of bills to curb AI harm
When artificial intelligence offers processing capacity for other formats, files produced by Claude will also receive watermark identification.
Understand the implementation of invisible watermarks in Claude’s texts
The watermarks present in the textual content created by Claude will be imperceptible to users and, as the company declares, will have no impact on the meaning, excellence or clarity of the AI’s responses.
As they are an intrinsic part of the text, these identifiers will remain with the content even after being copied and pasted in different locations, even being able to resist certain modifications.
Anthropic clarifies that “the watermark will be applied at the template level, meaning it will be present regardless of the product or Claude platform the text originates from.”
For files produced by Claude, source watermarks will adopt the C2PA open standard, widely recognized in the industry to indicate the provenance of content. The presence of this identification confirms that the file went through Claude’s processing.
See Also
Learn more: Regulation and advances in AI lead US bank to cut bitcoin price projection to US$82,000
Limitations and What Claude Watermarks Cannot Guarantee
However, the new watermarks implemented by Claude are not considered foolproof. Anthropic admits that the presence of identification does not conclusively attest that the content was completely created by AI.
The company emphasizes that “a detected watermark indicates that the content was processed by Claude, but is not entirely conclusive. The detection of a Claude mark signals that the content may have been processed by him. This does not, in itself, confirm the complete provenance of the content.” Therefore, Claude may not be the original creator of the material. An example of this is when AI is used to review, translate, summarize or convert files; the result may contain Claude’s stamp, even if the ideas, text or data come from another source.
Additionally, the watermark may remain on the content even after it has been modified subsequent to its processing by Claude. The company mentions that “the marked content can be modified, have excerpts extracted or be combined with other materials after being processed by Claude”.
Even if the watermark is not present, this does not serve as an absolute guarantee that the content was not created or processed by artificial intelligence. The indicator may not be displayed in the following situations:
- The material was generated by an older model of Claude;
- The text has been extensively edited, rewritten, translated or merged with other data;
- File metadata has been suppressed during document conversion, subsequent saves, screenshots, or other methods;
- The file or material was produced on a platform, tool, or format that does not yet support watermarking functionality.
Global applications of Claude watermarks driven by European law
The inclusion of watermarks aims to comply with the guidelines of the Code of Practice on Transparency of AI-Generated Content, which is part of the European Union AI Law. Although it is a regulation of regional scope, the measure will be applied globally in all locations where Claude is accessible.
Anthropic’s latest models, the Opus 5, released on July 24, 2026, and the Sonnet 5, released on June 30, 2026, do not incorporate the watermark by default. However, Anthropic guarantees that the feature will also be applied to these older versions of its AIs in the future.
