Claude’s Invisible Watermarks Now Silently Flag Every Piece of AI-Generated Text

Anthropic has added an invisible watermark to text generated by Claude, its AI assistant. The mark is a hidden signal built into the words Claude picks during generation. It doesn’t change what the text says or how it reads. Readers can’t see it. But Anthropic’s tools can detect it.

The watermark works by altering tiny random choices in how words are selected. It doesn’t use extra characters, hidden symbols, or special fonts. Instead, the signal lives in the text itself. A secret key or decoder is needed to verify it. That means ordinary people can’t spot it just by reading.

Anthropic says models launched on or after August 2, 2026 carry the watermark. The company has also said it’s working to add the feature to older models. The mark applies worldwide, not just in Europe. It covers Claude’s main platforms, including the API, Claude, Claude Code, Claude Cowork, and cloud partner integrations.

One important detail is what the watermark actually proves. It signals that Claude was involved in creating the text. It doesn’t prove Claude wrote everything from scratch. For example, if a person writes a paragraph and Claude makes a small edit, the signal may still show up. The mark flags Claude’s involvement, not full authorship.

The watermark signals Claude’s involvement in creating text, not that Claude authored it entirely.

The watermark can survive copying and pasting. It may also hold up through some editing. But Anthropic describes its durability as partial. Heavy changes to the text could break the signal. Translations produced by Claude also carry the watermark, so switching languages does not strip the signal.

Beyond text, Anthropic has added signed provenance metadata to certain file types. These include SVG, PNG, and JPG images. The file system uses C2PA-style metadata, which is separate from the text watermark. Together, the two layers aim to help identify AI-generated content.

There are limits. The watermark is invisible, so people can’t check for it on their own. Detection depends on Anthropic’s tools and keys. A public detection tool hasn’t shipped yet. That means for now, most people have no way to verify whether a piece of text came from Claude. The system marks the text, but checking it still requires Anthropic’s help. Non-compliance with the EU AI Act’s Article 50 mandate on marking AI-generated content can result in fines reaching €15 million or 3% of a company’s global annual turnover.