On August 2, Anthropic started marking AI-generated content from its newest Claude models. An invisible watermark in the text. Signed metadata on files. Worldwide, no opt-out.
The news broke wider this week, and the headlines are already sliding toward "everything you write with AI is now tracked."
That's not what's happening. But something real is — and for most content teams, it changes less than the headlines suggest. Here's what it actually touches.
What actually changed
Anthropic signed the EU AI Act's Code of Practice on marking AI content. Two mechanisms came with it.
Invisible text watermarking. Text from covered Claude models carries a machine-readable pattern. You can't see it. It survives copy-paste.
C2PA metadata on files. Files Claude produces carry signed provenance data recording that AI was involved. It's an open standard other AI companies use too.
Both apply globally — not just in the EU — across the apps, the API, Claude Code, and cloud deployments.
The part most headlines get wrong
The marking applies to models launched on or after August 2, 2026.
Models released before that date — which is most of what businesses run today — aren't marked yet. Anthropic says it's working to extend marking to older models during the EU transition period.
So "every Claude output is watermarked" isn't accurate. If your workflow runs on a model from before August, your output isn't carrying a watermark right now.
What the announcement really tells you is the direction. New models mark by default. Over time, the older ones follow. This isn't "everything changed on August 2." It's "the direction is now set."
What the mark actually proves
This is the part that will be most misunderstood, so it's worth being precise.
A watermark means content passed through a Claude model. It says nothing about how much of the thinking was human.
Paste your own fully-written draft into Claude to fix grammar? The output can carry the mark. Translate your own writing? Same. The mark says "AI processed this." It does not say "AI wrote this."
The reverse is also true. No mark doesn't prove a human wrote it. The text watermark weakens under paraphrasing and heavy editing. C2PA metadata doesn't survive a screenshot, a format conversion, or a re-save in a lot of tools.
So it cuts both ways. A detected watermark on your deliverable isn't evidence you didn't do the work. Clean-scanning content isn't evidence a human made it. Anyone treating these signals as proof — a client, an employer, a platform — is misusing them. That misuse is where the first real problems will come from.
How marketers actually use Claude — and what the mark touches
This stays abstract until you map it onto real work. Here's how AI actually shows up in most marketing operations, and how differently the change lands on each.
Editing and polish. A human writes it, Claude tightens it — grammar, length, tone. All the thinking is human. But the marked-up text comes back out of the model, so it can carry the mark. This is the case that most needs the processing-versus-authorship distinction.
First drafts from a human brief. You supply the outline, positioning, key points, sources. Claude produces a draft. You rewrite and fact-check. Here the mark is arguably fair — AI genuinely generated the prose.
Volume variations. Twenty ad headlines, ten subject lines, five hooks. Almost pure generation, and one of the highest-value AI uses in marketing. Few people would claim otherwise.
Repurposing. Webinar transcript into a blog post. Blog post into a carousel. Case study into an email sequence. The content is entirely yours; AI is doing format conversion.
Research and synthesis. Summarizing competitor pages, pulling themes from customer interviews. Often never published — it informs a human who writes something original.
Translation. Your approved copy in another language. Fully human source, model-processed output.
Notice the pattern. In most of these — editing, repurposing, translation, research — the human supplied the substance and AI handled the mechanics. Only first drafts and volume variations involve AI producing the ideas.
A provenance mark looks identical across all of them.
That's why blanket claims fail in both directions. "We don't use AI" is untrue for nearly every modern content team. "This was AI-generated" is misleading when a model only translated your own writing. The accurate statement is always about where in the process AI sits.
Can you check your own content?
Two different answers.
For files, yes. C2PA is an open standard, and the Content Authenticity Initiative runs a free inspection tool at contentcredentials.org. Upload a file, see any signed provenance data attached. Worth running a few of your own recent deliverables through it, just so you know what you're shipping.
For text, not yet. Anthropic hasn't released a public detection tool. The watermark is built for detection systems, not human eyes. You can't check a block of text yourself — and neither can your clients.
One caution as "AI detectors" market themselves off this news: a watermark check and an AI-writing-style detector are completely different things. Style detectors guess from word patterns and are famously unreliable — they flag human writing as AI and miss AI writing constantly. A watermark is a deliberate embedded signal. If a tool can't tell you which one it's reading, don't trust it.
What to do
Know what your stack runs on. Find out which models your tools actually use. Plenty of marketing platforms run AI under the hood without saying which model. Whether your output carries a mark depends on that — and it'll change as vendors upgrade.
Decide your disclosure posture before someone asks. The uncomfortable moment isn't the watermark. It's a client running a detection tool on your deliverable and asking a question you haven't thought about. Teams that use AI and say so plainly have nothing to navigate. Teams whose positioning implies untouched human authorship now have a gap between the claim and the reality. Close it on your own terms.
Revisit "100% human-written" claims. If that's in your proposals, decide what you actually mean. Human-directed? Human-edited? No AI at any step? Write down the honest version now. Vague authorship claims age badly in a world with provenance metadata.
Don't panic-rebuild your workflow. Nothing here changes what your content is or how it performs. Search engines rank on quality and usefulness, not provenance marks. Good content with AI in the process was fine before August 2 and is fine after. This is about transparency, not penalty.
The bigger picture
Anthropic is the first major lab to implement the EU framework at this scale, and the direction is clear. Other providers signed the same code. C2PA adoption is spreading. Provenance signals are becoming infrastructure — the way HTTPS went from novelty to default.
For content teams that's less a threat than a clarifying force. The era of ambiguity about AI involvement is ending.
The teams most comfortable with what comes next are the ones whose value never depended on that ambiguity — where the strategy, the judgment, and the quality are the product, and the tools are just tools. If that's your operation, this asks nothing of you except accuracy about your own process.
That was always worth having.
