Anthropic Adds Invisible Watermarks to Claude-Generated Content

Anthropic has begun adding machine-readable marks in content produced by its Claude AI models, making it easier to identify AI-generated text and files. This move is detailed in an updated Claude Help Center article, aligns with the company’s commitments under the European Union’s AI Act.

Claude models launched in the EU on or after August 2, 2026, support the marking system from day one. Anthropic is applying it worldwide across supported models, not just in Europe. The company is also working to add marking support to earlier models during a transition period allowed under the law.

How the Marking Works

Anthropic uses two complementary techniques:

• Embedded text watermarks: When a supported Claude model generates text, it weaves an imperceptible watermark directly into the output. The mark is invisible to human readers and does not change the original meaning, quality, or readability of the response. Because it is part of the text itself, it can travel with content when copied and pasted and may survive some editing. The watermark is applied at the model level, so it appears regardless of whether users access Claude through the website, API, Claude Code, Claude Cowork, Claude Tag, or cloud partners such as AWS, Google Cloud, and Microsoft Foundry.

• Signed provenance metadata for files: For supported file types such as .svg, .png, and .jpg, Claude attaches digitally signed metadata following the Coalition for Content Provenance and Authenticity (C2PA) open standard. This can indicate that a file was generated or processed by Claude and help detect tampering.

Anthropic emphasizes that detecting a Claude mark does not prove the company was the original author. Users frequently employ Claude to proofread, translate, summarize, or convert existing human-written material, so the output can carry a mark even when the underlying ideas or text originated elsewhere.

The company plans to provide details and tools so users and third parties can detect the marks, as required by the EU Code of Practice. It notes that marks may not always be detectable such as, after heavy editing, with very short text, or if metadata is stripped.

The initiative stems from Anthropic signing the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content, a voluntary framework intended to make AI-generated material easier to identify.

The change arrives as regulators and platforms worldwide push for greater transparency around AI-generated content, often called “AI slop” in online discussions. Critics have long noted that text is harder to flag than images or video, and Anthropic’s model-level approach aims to address that by making the signal harder to remove through simple copying.