Artificial intelligence company Anthropic has revealed comprehensive details regarding its new watermarking technology for Claude, designed to embed invisible, machine-readable identifiers directly into AI-generated text and media files. The move comes as part of the company’s efforts to comply with the European Union’s AI Act transparency mandates, following the EU Code of Practice on Transparency of AI-Generated Content that Anthropic and other major AI providers signed in July 2026.
The watermarking system is being applied globally at launch, as Anthropic says it does not yet have a durable way to scope the feature by region.
Explaining the mechanics behind the feature, Anthropic noted that the text watermarking does not rely on hidden Unicode characters or visual tags. Instead, the model incorporates a subtle statistical pattern into its natural word selection process during generation. While the text remains indistinguishable to human readers and does not alter output quality or token consumption, automated detection tools equipped with a decoding key will be able to determine the probability that the content was generated or processed by Claude.
In addition to text, Claude will attach cryptographically signed provenance metadata following Coalition for Content Provenance and Authenticity (C2PA) standards for generated image and vector files.
Anthropic clarified that while light editing likely won’t remove the watermark, a complete rewrite where every word is replaced will. A detected mark indicates that Claude processed the text at some stage, rather than proving complete automated authorship — accounting for workflows such as proofreading and translation. All models released on or after August 2, 2026, include the feature at launch, with legacy models slated to be updated over the coming months.