Anthropic Adds Invisible Watermarks to Claude-Generated Text
The AI company is introducing machine-readable watermarks in Claude’s generated text, along with signed provenance metadata for supported files, as part of a broader push for transparency around AI-created content.
Anthropic has introduced a new system that embeds an imperceptible watermark into text generated by supported Claude models. The watermark is designed to be invisible to readers and should not change the meaning or readability of the content. Unlike a visible label, the mark is woven into the generated text itself, allowing specialized detection systems to identify content produced by Claude.
The technology is being introduced as AI-generated writing becomes increasingly common across education, publishing, business and online media. Anthropic says the watermark is designed to remain present when text is copied and pasted and may survive some forms of editing. However, heavy rewriting, translation or combining AI-generated material with other text can affect whether the watermark remains detectable.
The move is also connected to new European transparency requirements. The European Union’s AI Act introduces transparency obligations for AI-generated or manipulated content, with key requirements taking effect on August 2, 2026. Anthropic’s approach is intended to help establish clearer provenance for content produced by its models and make it easier for people and organizations to distinguish AI-generated material from human-created work.
Anthropic is extending the provenance approach beyond text. Supported files generated by Claude can include digitally signed metadata based on the C2PA standard, providing information about their origin and helping users verify whether provenance information has been altered. The company also says it is working toward enabling third parties to detect the watermarks, which could make the technology more useful for publishers, educators and other organizations.
The development represents a significant shift in how AI companies approach content authenticity. Watermarking could make AI-generated material easier to trace, although it is not a perfect solution because extensive transformation can weaken or remove detectable signals. Still, Anthropic’s decision puts greater emphasis on provenance and transparency as AI-generated content becomes a larger part of everyday digital communication. The move could also encourage other AI developers to adopt similar systems as governments and organizations seek better ways to understand where digital content comes from.
