Everything Claude Writes Now Carries an Invisible Watermark

If you've used Claude to draft anything published under your name since early August, that text is carrying a signature you can't see and can't remove by lightly editing it. Anthropic confirmed on August 11 that every Claude model released after August 2 embeds an invisible, machine-readable watermark into its generated text — and the detail that matters for your work is that it's built to survive copy-paste and moderate rewriting.

What happened

The watermark uses an approach based on Google DeepMind's SynthID-Text: it subtly biases word choices in a statistically detectable pattern, applied globally across the API, claude.ai, Claude Code, Claude Cowork, and cloud partner deployments — not just in the EU. The trigger is the EU AI Act's Article 50 transparency requirement, which took effect August 2, but Anthropic didn't scope the rollout to European users. Developer Sophie Alpert's stated rationale, amplified by Simon Willison on August 11: "There are no lossless transformations of natural-language text — every rewrite changes meaning," meaning a watermark that only survived unedited output wouldn't be worth building.

Not everyone thinks this was the right call. Ben Thompson's August 12 Stratechery analysis called mandatory watermarking "a terrible idea," comparing it to "insisting that a ballpoint pen advertise itself as the author." The disagreement is philosophical, but the practical effect lands on you regardless of who's right: your AI-assisted output is now identifiable as AI-assisted, by design, and other providers will likely follow since the EU AI Act applies to all of them.

The skill implication

There's a legal wrinkle underneath the technical one. The U.S. Copyright Office's position is that purely AI-generated material needs "sufficient human authorship" to qualify for copyright protection. If your deliverable is machine-identifiable as AI-generated, your ownership claim over it depends on how much you actually contributed — and now there's a way to check.

This changes what "I edited it" needs to mean. Three questions worth running before anything AI-assisted goes out under your name:

Can you explain every claim in it without opening the chat again? If not, you haven't revised it — you've forwarded it.

Did you change the argument, or just the phrasing? A watermark survives phrasing changes. Ownership arguably requires more than that.

Would you stand behind this if someone could prove which parts came from the model? That's no longer hypothetical.

Two people, two exposures

Dana runs marketing at a 40-person SaaS company and used Claude to draft a client case study — a first pass, then two rounds of her own edits pulling in specifics from the actual client call. Her exposure here is disclosure, not detection: nothing she published was factually wrong, but if a client or a competitor ever ran the piece through a detector and asked "did AI write this," her honest answer needs to match what actually happened, and "I drafted from a Claude outline, then rewrote the structure and added the client's real numbers" is a very different answer than "I wrote this."

Marcus freelances as a business writer for trade publications and is more exposed on the ownership side. A piece he drafts largely with Claude and files with light editing is now provably closer to "AI-generated" than "human-authored" under the Copyright Office's framework — which matters directly for a freelancer, since his ability to license and resell that work depends on being able to claim it's his.

Both of them can protect themselves the same way: treat "I used AI to draft this" and "I own this" as two separate claims, and make sure the second one is actually true before you publish, not just assumed.

The takeaway

The watermark doesn't make AI-assisted writing wrong — it makes the gap between "I used AI" and "I did the work" checkable for the first time, so the review-and-revision step you might have been skipping now has a receipt.