Hey — Chris here.
Anthropic confirmed today that it will watermark the text its models generate — across Claude, the API, Claude Code, Cowork, and Tag. The trigger is the EU AI Act's new transparency rules (in force since August 2), but the marking applies worldwide, wherever Claude runs.
The kitchen-table version: invisible ink woven into the words themselves. The sentence reads exactly the same to you, but a scanner can see the signature. And because it's embedded in how the model picks its words — not a tag stapled on afterward — it survives copy-paste and may survive some editing too.
What's actually known
New models released after August 2 will mark text and files from day one (files use C2PA, the same content-credentials standard cameras and image tools use). Older models don't mark yet, but Anthropic says they'll get it during the AI Act's transition period.
And it's not just Anthropic. Google, Meta, Microsoft, and OpenAI signed the same EU code. Suno is watermarking generated music. Substack just partnered with a detection company called Pangram to flag AI-generated newsletters — yes, newsletters, the thing you're reading right now.
What's genuinely unknown
Three things nobody has answered yet: how much editing removes the watermark (assume durability), whether generated code carries it (Claude Code is on the list, and code is text), and who gets access to the detector. I'm not going to pretend to know. I'll report back when someone does.
My take
I draft this newsletter and my blog with my AI crew, and I've said so from the start. So for me, a signature that says "a model helped write this" confirms something I already tell you. It costs me nothing.
It costs a lot if your whole model depends on AI text passing as fully human. That business is now on a countdown — a slow one, maybe, since older models don't mark yet. But every platform that adds scanning shortens it, and Substack already moved.
The move this week isn't technical, it's a sentence: put your AI disclosure on your about page before a scanner puts it there for you. Disclosure you volunteer reads as honesty. Disclosure that gets detected reads as getting caught.
I broke down the full known/unknown ledger, the SEO question, and my do/don't list in the full post:
Try it yourself: write your one-sentence disclosure right now — "posts on this site are drafted with AI tools and edited by me" — and put it where readers can find it. Took me under a minute, and it's the only prep this announcement actually requires.
— Chris