Anthropic Is Now Watermarking Every Word Claude Writes. Here's What That Means If You Resell AI Content.

Follow-up to: 2026-08-03, the EU AI Act transparency piece. That post covered the regulation itself: who it applies to, what it requires, what the fines look like. This one covers a specific company's response to it, and it's genuinely new information: Anthropic's watermarking mechanics, who else signed onto the same commitment, and what the mark actually catches.

On August 11, 2026, Anthropic confirmed something I'd been waiting to see spelled out since I wrote about the AI Act's transparency rules going live nine days earlier: it's now weaving an invisible watermark into every piece of text a supported Claude model generates. Not just Claude.ai output. Not just for EU accounts. Every supported model, every product surface, worldwide, according to Anthropic's own support documentation. If you've been quietly reselling Claude-drafted posts, ghostwriting under your own byline, or repackaging AI output as human work without saying so, there's now a standing technical mechanism, imperfect as it is, that can flag it.

What Anthropic actually shipped

Anthropic signed Article 50(2) of the EU AI Act's Code of Practice on Transparency of AI-Generated Content, and its updated Help Center article lays out the mechanics directly. Claude models launched on or after August 2, 2026 support marking from day one: an embedded watermark in generated text, plus signed provenance metadata (following the C2PA standard) in generated files like SVGs, PNGs, and JPGs. Anthropic's own description of the text watermark is worth quoting exactly, because it answers the "how imperceptible" question a lot of people are asking: it "weaves an imperceptible watermark directly into the text itself," one that "doesn't change the meaning, quality, or readability" of the response, and because it's part of the text rather than a wrapper around it, "it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing."

The mark is applied at the model level, so it shows up regardless of which surface you're using: Claude.ai, Claude Code, Claude Cowork, Claude Tag, the API, and Claude accessed through AWS Bedrock, Google Cloud, or Microsoft Foundry. And it's worldwide by design, not by accident. Anthropic states plainly that marking "will apply to output from supported models wherever Claude is offered, worldwide." I went looking for a region-gated version of this and there isn't one. The likeliest explanation, and the one that fits how every other frontier lab ships product, is that building a separate behavior tree for EU-origin requests is more engineering overhead than just shipping one global default. Compliance became the reason to build it, not the boundary of where it applies.

One real gap: this only covers models launched on or after August 2. Anthropic calls out a "transition period" for anything released before that date and says marking support for older models is still in progress. If you're on an older model, there's no watermark yet, at least not officially.

The rest of the industry signed the same code

This isn't just an Anthropic thing, and I want to be precise about who else is on the list rather than repeat a shorthand version of it. The European Commission reported that about 190 organizations had signed the Code of Practice on Transparency of AI-Generated Content by the end of July 2026, ahead of the August 2 enforcement date. The Code splits into two sections: one covering the marking obligation itself (Section 1), one covering deployer-side disclosure and labeling (Section 2). Section 1 has 82 signatories, and the Commission's own release names the "well-established and prominent AI companies" on it explicitly: Aleph Alpha, Anthropic, Black Forest Labs, Cohere, Google, Meta, Microsoft, Mistral, OpenAI, and Synthesia. That's not a handful of majors making a symbolic gesture. That's basically the entire frontier-model tier plus the leading open-weight and enterprise players, all agreeing to the same marking commitment Anthropic just shipped.

The Commission also describes the Code as having been "assessed as adequate" by both the Commission and the AI Board, which gives signatories what it calls a "streamlined, predictable and legally certain pathway" to demonstrate Article 50 compliance, regardless of where the company is incorporated. That's the same territorial-scope point from the earlier piece, just showing up from the compliance-mechanism side this time: it doesn't matter where Anthropic, OpenAI, or Mistral are headquartered, what matters is that their output reaches someone in the EU.

What this means if you resell or ghostwrite AI content

Here's the part that actually matters for a solo operator, and it has almost nothing to do with EU regulators showing up at your door. If you sell "human-written" content, run a ghostwriting service where the client contract assumes you're the one typing, or repackage Claude drafts as your own newsletter or blog posts without disclosure, the risk that's grown in the last two weeks isn't a fine. It's that the tooling ecosystem around detecting this stuff is getting real fast. Substack partnered with the detection company Pangram in July specifically to flag AI-generated newsletters, and Substack's CEO Chris Best has been using the term "Claudefishing" for writers who pass off AI output as their own. Anthropic's watermark adds another layer to that same trend: a mark embedded at the model level, present across every Claude product, that third parties will eventually be able to check for once Anthropic ships the detection tooling it says is coming.

Practically, that means the exposure most solo operators should worry about is contractual and reputational, not regulatory. A client who hired you assuming human-drafted work has more immediate leverage than the AI Office does over a one-person shop. The Reddit reaction to Anthropic's announcement, covered by TechCrunch the following day, split how you'd expect: a minority called the watermark invasive, a larger share pushed back with some version of "there's no good argument against this unless you're planning to lie to people." That's roughly where I land too.

The honest take

The framing I'd push back on hardest, including in my own headline above, is treating this watermark as some kind of foolproof surveillance layer. Anthropic's own support article includes a Limitations section that undercuts the "gotcha" story on its own terms: a detected mark isn't conclusive proof of authorship, since Claude is often used to proofread or summarize someone else's writing, and the mark can disappear entirely if the text is "heavily edited, paraphrased, translated, or mixed into other writing," or if the passage is too short to carry a reliable signal. That's Anthropic saying this out loud, not a critic.

That admission lines up with independent research on how fragile text watermarking is in general. ETH Zürich researchers Robin Staab and Nikola Jovanović published work in 2024 (covered by MIT Technology Review) showing they could strip the common "green list/red list" style of LLM watermark about 85% of the time and spoof one onto human-written text about 80% of the time, just by querying the watermarked model repeatedly and reverse-engineering its rules. Soheil Feizi's group at the University of Maryland found similar fragility independently. Anthropic hasn't published the technical details of its own watermarking scheme, so I can't say with certainty it uses the exact same green-list approach those attacks target. But the general finding, that invisible text watermarks are a genuinely unsolved research problem and not a settled one, holds regardless of the specific algorithm, and Anthropic's own limitations language is consistent with it.

What I'd actually do: if you're reselling or ghostwriting Claude output without disclosure, don't treat "the watermark probably won't survive my edit pass" as a green light to keep doing it quietly. Do one of two things instead. Either disclose that the work is AI-assisted, which costs you nothing with most clients and avoids the entire question, or actually do enough substantive rewriting that what you're selling isn't Claude's sentence structure with a coat of paint, in which case the watermark question becomes moot because you did the work the client is paying for. The honest counter-take is that for most solo operators, this specific mechanism is unlikely to be what catches you: a real paraphrase pass, a translation round-trip, or even normal heavy editing will probably defeat it, by Anthropic's own account and the independent research above. Treat this as one more sign that AI-content detection tooling is getting more common and more embedded, not as a single mechanism you need to specifically outsmart.

Author

Lukas

@lukcombinator

Sources

• How Claude marks AI-generated content (Anthropic Help Center)
• Anthropic says it will watermark text generated by its AI models
• Some Claude users are mad that Anthropic's new watermarks will catch them using it at their jobs, classes
• Strong backing for the Code of Practice on Transparency of AI-generated Content (European Commission)
• It's easy to tamper with watermarks from AI-generated text (MIT Technology Review)