Anthropic, in an official weblog publish has introduced that Claude can embed invisible watermarks in AI-generated textual content, making a hidden path that may survive copy-pasting.
So, should you ask Claude to jot down one thing for you, there might now be a hidden marker sitting inside that textual content.
You will not see it. It will not change how the textual content reads. However Anthropic mentioned supported Claude fashions can embed an imperceptible, machine-readable watermark instantly into AI-generated textual content.
And since the watermark is a part of the textual content itself, it might journey with the textual content when it’s copied and pasted elsewhere, and should survive some modifying.
The transfer is a part of Anthropic’s commitments below the EU AI Act’s Code of Apply on Transparency of AI-Generated Content material. Claude fashions launched within the European Union on or after August 2, 2026 will assist machine-readable marking from launch. Anthropic says the marking will even apply worldwide to supported fashions.
The system covers Claude’s API in addition to merchandise together with Claude, Claude Code, Claude Cowork and Claude Tag.
However textual content is not the one factor Claude can mark. For supported recordsdata akin to PNG, JPG and SVG, Anthropic mentioned Claude will connect digitally signed provenance metadata primarily based on the C2PA normal. This will point out {that a} file was processed by Claude and assist detect whether or not the metadata has been tampered with.
There is a crucial rider although. Anthropic mentioned a detected watermark shouldn’t be conclusive proof of the place content material got here from.
Somebody could have written the unique materials themselves after which used Claude to proofread, translate, summarise or convert it. The ensuing content material can nonetheless carry a Claude mark.
Content material can even change after Claude processes it. It might be edited, excerpted or mixed with different materials.
And the absence of a detectable mark does not essentially imply AI wasn’t concerned.
Anthropic mentioned a Claude-generated passage could not carry a detectable mark if it got here from an older mannequin, was closely edited, paraphrased or translated, is simply too brief to offer a dependable sign, or has been blended into different writing.
For recordsdata, provenance metadata can even disappear by means of format conversion, re-saving or screenshots.
Anthropic remains to be engaged on the instruments that may permit customers and third events to detect Claude’s watermarks and provenance metadata. The corporate will publish extra technical particulars on how detection works.
So this is not fairly an AI lie detector. It is extra like a hidden digital path, designed to provide customers and platforms one other sign about whether or not content material could have been processed by Claude.
