Anthropic Announces Claude's Invisible Text Watermarking Scheme, Invisible Marks Becoming Default Output
Anthropic announced that Claude will embed invisible watermarks in generated text as part of its content marking scheme.
The watermark is directly woven into the text at the model level, without altering the meaning, quality, or readability of the responses, and will be carried by all Claude products regardless of the output.
The watermark will transfer with text when copied and pasted, and may still be retained after a certain degree of editing.
This initiative applies to new Claude models released in the EU on or after August 2, 2026, and will be effective globally, covering all product lines including API, Claude's official website, and Claude Code.
Additionally, files generated by Claude in SVG, PNG, JPG formats will include digital signature provenance metadata compliant with C2PA standards.
From a market mechanism perspective, mandatory machine-readable markings enhance the recognizability of AI-generated content, compliance requirements drive model providers to invest in watermark technology, and reliance on unmarked content usage scenarios is under pressure, while event-driven transparency standards accelerate unification.
Anthropic clearly states that detecting a mark does not necessarily prove Claude is the original author, as users may use it for proofreading, translation, or summarization.
Source: Public Information
ABAB AI Insight
Anthropic's deployment of text-level invisible watermarks directly responds to the EU AI Act Article 50's requirements for transparency in generated content, opting for global applicability rather than limiting it to the EU.
The capital and technical pathway is reflected in embedding the watermark in the model output layer rather than post-processing, motivated by ensuring that the mark propagates with the text and is difficult to easily strip away, while also forming a dual provenance system with C2PA file metadata.
This is similar to OpenAI's previous development of text watermarks, which have not been fully launched, and the practices of image generation tools that have adopted C2PA, currently transitioning from voluntary disclosure by large model vendors to mandatory machine-readable markings.
Essentially, this is a regulatory change, with the mechanism being that once generated content scales, detectability becomes a necessary infrastructure to maintain the credibility of the information environment.
ABAB News · Cognitive Law
Invisible marks are becoming the default output
Model-level watermarks are harder to strip than post-processing
Transparency requirements are reshaping the fundamental attributes of generated content