Everything Claude Writes Now Carries a Hidden Mark, Everywhere in the World
Anthropic will embed invisible watermarks in Claude's text and sign generated files with C2PA metadata. The company is unusually clear about what the marks cannot tell you.
Anthropic has signed the EU AI Act Code of Practice on transparency, and the practical result lands today: Claude models launching from 2 August 2026 onward will stamp an invisible watermark into every piece of text they write. Generated files such as PNG, JPG and SVG images get signed provenance metadata instead. The rule was written for Europe, but Anthropic is applying it worldwide, across the API, Claude, Claude Code, Claude Cowork and Claude Tag.
A watermark here does not mean a visible logo. For text, the model nudges its own word choices in a pattern that a detector can recognise but a reader cannot, without changing meaning or readability, according to the company. It survives copy and paste and, in Anthropic’s careful phrasing, “may persist through some editing.” For files, the approach is different: a digital signature following the open C2PA standard, developed by the Coalition for Content Provenance and Authenticity, records that Claude touched the file and reveals later tampering. Older models get a transition period, and Anthropic says it is working on retrofitting them. Verification tools for the public are promised, without a date.
The honest part is the list of things this does not prove. A watermark does not mean Claude wrote the ideas, because people constantly use it to proofread, translate or shorten their own writing. Equally, no watermark proves nothing: the text may come from an older model, may have been heavily rewritten, may be too short to detect, or the metadata may have been stripped by a format conversion or a screenshot. Anthropic is not alone in trying. Google DeepMind open sourced its SynthID system for the same job, and OpenAI has sat on a text detector it says is 99.9 percent accurate for about two years without releasing it, partly because translation defeats it and partly because false accusations do real damage.
That damage is the reason this matters beyond compliance paperwork. Detection tools have already been used to accuse students of cheating on evidence nobody can inspect. A watermark that the maker itself publishes and lets others verify is a different kind of evidence: you can check what it is based on. It could also cost Anthropic customers, since Claude is popular with students and knowledge workers precisely because its prose reads naturally.
What this means for you: if you use Claude for work you sign your name to, assume the text is now marked, and assume the mark may travel with it. That is not a reason to stop using it, since editing and rewriting weaken the signal and the mark says nothing about who had the ideas. But if you are a teacher, an editor or a hiring manager, treat any detector result as one weak signal among many, never as proof. And if you publish AI assisted work in the EU, remember that Article 50 puts the disclosure duty on you, not on the model.
Sources
Source: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
Nvidia Will Cover Part of the Loss If Its Own Chips Lose Value. That Is How You Raise 500 Billion Dollars.
Six of the biggest finance houses want to fund AI data centres. To make the numbers work, Nvidia is guaranteeing up to a quarter of what its hardware will be worth at the end.
