If you use Claude for anything you write, there’s a new detail worth knowing. As of August 11, 2026, Anthropic confirmed that Claude now embeds an invisible watermark into the text it generates. No visible mark, no formatting change, nothing you’d notice on the page. It’s baked into the writing itself.
Here’s what’s going on and why it matters.
Why This Is Happening
The short answer is regulation. The EU AI Act’s Transparency Code took effect on August 2, 2026, and it requires AI companies to mark content their systems generate so it can be identified later. Anthropic could have limited this to European users. Instead, they rolled it out everywhere, for every Claude product: the API, claude.ai, Claude Code, Claude Cowork, and Claude Tag.
How It Works
When a Claude model generates text, it weaves a statistical pattern into the writing. It’s not a stamp or a tag you can spot. Anthropic says it doesn’t change the meaning, quality, or readability of the response. The watermark rides along with the text even when you copy and paste it somewhere else, and it can survive some editing.
This applies to models launched on or after August 2, 2026, which covers current models like Claude Sonnet 4.6 and Claude Haiku 4.5. Older models don’t have it yet, though Anthropic says support is coming for those too.
Files get a related but separate treatment. Images and other generated files carry signed provenance metadata using the C2PA standard, an open format several companies are adopting together.
The Part Worth Understanding
Here’s the piece people tend to miss. The watermark doesn’t tell you Claude wrote something. It tells you Claude touched it. If you write a document yourself and just ask Claude to proofread it or clean up the formatting, that document still comes back marked. Processed and authored aren’t the same thing, but the watermark can’t tell them apart.
It also isn’t bulletproof. Heavy editing, paraphrasing, translation, or mixing the text with other writing can weaken or erase the signal. Short passages might not carry enough text for a reliable read either.
Where This Fits in the Bigger Picture
Anthropic isn’t the first to try this. Google has been watermarking Gemini output with its own system, called SynthID, since 2023, and it’s now built into Search, Chrome, and a standalone detector tool. The two systems aren’t the same, though, and right now they don’t talk to each other. Google’s detection tools recognize SynthID and C2PA metadata specifically. Claude’s text watermark is its own approach, so it isn’t something Google’s existing tools would catch on their own.
Anthropic has also said a public detection tool is on the way, something people could use to check text themselves. That’s not available yet.
The Bottom Line
For now, nobody’s reading this watermark with the naked eye, and there’s no widely available tool checking for it either. That’s likely to change as detection tools roll out and more companies adopt shared standards. If you use AI tools for your writing, this is a good moment to know it’s happening, even if it’s invisible today.