Claude Will Now Secretly Watermark All AI-Generated Text and Images

Claude is getting a new feature nobody asked for, and most users won’t even notice it’s there. Anthropic confirmed this week that its AI models will start embedding invisible watermarks into generated text, plus digital tags on images.

This isn’t optional, and it is not something you will see. The mark sits inside the text itself, invisible, and according to Anthropic, it doesn’t change what Claude writes or how it reads.

New models launched from August 2 onward carry this by default. That covers the regular Claude app, the API, Claude Code, Cowork, and Claude Tag too. Basically anywhere Claude generates something, the mark goes with it.

Copy the text somewhere else, paste it into a document, an email, whatever, the watermark is supposed to travel along with it. Anthropic says it can even survive some editing, though not all.

Images work differently. Instead of a hidden signal buried in pixels, Claude attaches a metadata file using something called C2PA, a standard already used by companies like Adobe and Google. This covers JPG, PNG and SVG files, and the metadata records which model made the image and when.

Why now? The EU AI Act is why. Since August 2, the law requires AI tools that generate synthetic content, deepfakes especially, to clearly label that content as machine-made. Anthropic signed onto this requirement.

Here’s the part that surprised some people. Even though this is an EU rule, Anthropic isn’t limiting it to European users. The watermark applies everywhere Claude runs, AWS, Google Cloud, Microsoft Foundry included, not just claude.ai.

Older Claude models don’t have this yet. Anthropic says it’s still working on retrofitting them during the transition period the EU law allows.

The company has been upfront that this isn’t a perfect system. A detected watermark doesn’t prove Claude wrote something entirely on its own, it could’ve just proofread or translated existing text. And no watermark doesn’t mean a human wrote it either. Heavy edits, short snippets, format changes, all of that can wipe the signal clean.

Anthropic isn’t the only one doing this. Google, Meta, Microsoft, OpenAI, Black Forest Labs and Synthesia have all signed onto similar commitments under the same EU code.

Reaction online hasn’t been entirely positive. Since the announcement, plenty of users have pushed back at the idea that their AI-assisted writing could quietly carry a tag marking it as machine-generated, arguing it feels less like transparency and more like surveillance.