Anthropic is putting a hidden AI watermark in the text Claude generates, and it will apply everywhere Claude is offered, worldwide, not just in the EU. The company laid out the plan in an update to its support pages, and it’s a bigger deal for regular users than the dry phrasing suggests.
The change comes as Anthropic signs the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. But the marking won’t be limited to European users. New Claude models launched on or after August 2, 2026 will support machine-readable marking from day one, and Anthropic says existing models will gain the feature during the transition period.
How the AI watermark actually works
The watermark is woven directly into the generated text. You won’t see it, and it doesn’t change the meaning, quality, or readability of a response. Because the mark is part of the text, it travels with the text when you copy and paste it elsewhere, and it may survive some editing.
The marking is applied at the model level. That means it’s present no matter which Claude surface the text comes from: Claude, the Claude API, Claude Code, Claude Cowork, and Claude Tag. Supported models accessed through AWS, Google Cloud, or Microsoft Foundry will embed the watermark too, where supported.
For files, Claude attaches signed provenance metadata to supported types like PNG, JPG, and SVG. That metadata follows the C2PA open standard and records that the file was processed by Claude.
The design is deliberately redundant. Text gets the invisible mark, files get the metadata, and both signals point back to the same fact: this content passed through Claude at some point. That redundancy matters because no single technique survives every path. Copy-paste kills some metadata, screenshots kill other metadata, and the text watermark is the part that sticks around through normal use.
Anthropic is treating marking as a property of the model rather than a feature of one app, and that’s the only way the system stays consistent across every integration. If you call Claude through an API, the mark is already in the reply before your code ever touches it.
What the AI watermark proves, and what it doesn’t
This is the part worth reading twice. A detected mark is not proof that Claude originally wrote the text. It signals that the content may have been processed by Claude. People use Claude to proofread, translate, summarize, or reformat their own work all the time, so the mark doesn’t tell you who the author is.
The reverse is true too. A missing mark doesn’t mean the text wasn’t AI-generated. Heavy editing, paraphrasing, translation, short passages, or combining Claude output with other writing can make the watermark undetectable.
So think of the AI watermark as a soft signal about provenance, not a hard verdict on authorship. Anthropic hasn’t yet said how users or third parties will detect the marks, only that detection mechanisms will come in future documentation. Until that detection tooling ships, the marks are one-way: Claude writes them, and nobody outside Anthropic can reliably read them yet.
That open question is the honest gap in this announcement. Embedding a mark is the easy half. Giving the public a trustworthy way to check for it is the hard half, and it’s the half that decides whether this becomes a real transparency tool or a compliance checkbox.
Why this matters for anyone who writes with AI
For the growing number of people who lean on Claude for work, the takeaway is that AI output is quietly becoming tagged, and the tag can outlive the chat window.
Anthropic has had a tough stretch, from models hacking companies during security evals to the broader trust questions swirling around agents. Watermarking is the company’s attempt to answer a different part of that trust problem: knowing where content came from.
AI-generated content has already burned people this year, and the industry keeps reaching for transparency tools as a fix. Anthropic isn’t first. Google and Meta have also signed the EU transparency code, and Suno said it would start watermarking songs.
The practical read is mixed. If you publish AI-assisted work, a watermark you can’t see shouldn’t change what you disclose. If you’re worried about AI text being passed off as human, an AI watermark is a step in the right direction, but it’s not the definitive test a lot of people want it to be.
Bottom line
Anthropic is betting that knowing the origin of text is worth building into every response. That’s a meaningful shift for the whole chatbot category, because the default has always been silence about provenance.
For now the mark is invisible and easy to forget. But it’s another sign that the industry is moving past asking whether text was AI-written and toward embedding the answer in the text itself. I’ll be watching how the detection side lands, because that’s where the real test happens.
