On August 2, Anthropic began marking AI-generated content from its newest Claude models. Text carries an invisible watermark. Supported file exports carry signed provenance metadata. It applies worldwide, and there's no opt-out.
The coverage since has drifted toward "everything you write with AI is now tracked," which overstates both what is marked and what a mark proves. Here's what the change actually touches.
What shipped
Anthropic signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, and two mechanisms came with it.
Text from covered Claude models carries a machine-readable pattern woven into the wording. You can't see it, and Anthropic says it doesn't change the meaning, quality, or readability of the output. Because it sits inside the words rather than alongside them, it survives copy-paste and can survive a degree of editing.
Supported file exports — .svg, .png, and .jpg — carry signed provenance metadata built on C2PA, the open standard Adobe, Google, and OpenAI already use. That metadata records that Claude was involved and allows later tampering to be detected.
Both apply globally rather than only in the EU, across Claude Platform, Claude, Claude Code, Claude Cowork, and Claude Tag. Anthropic applies it worldwide because it doesn't yet have a reliable way to limit the measure by region. Text watermarks also persist when supported models are reached through AWS, Google Cloud, or Microsoft Foundry, though signed file metadata may not be available on every platform depending on what each one supports.
Which models are actually marked
Marking applies to models launched on or after August 2, 2026.
Models released before that date — most of what businesses run today — fall inside a transition period written into the EU rules, reported to run four months. Anthropic describes retrofitting work on older models as under way, with its support page to be updated as it becomes available.
So if your workflow runs on a pre-August model, your output isn't carrying a text watermark right now. What the announcement establishes is direction: new models mark by default, older ones follow.
What a mark proves
This is the part most likely to be misused, so it's worth being exact.
A watermark indicates content passed through a Claude model. It says nothing about how much of the thinking was human. Paste your own finished draft in to fix grammar and the output can carry the mark. Translate your own writing and the same applies. The signal is that AI processed the text, not that AI wrote it.
Anthropic is direct about the limits in its own documentation, describing a detected mark as a signal that content was processed by Claude while stating it is "not fully conclusive." Marks can be absent after editing or paraphrasing. C2PA metadata doesn't survive a screenshot, a format conversion, or a re-save in plenty of tools.
It cuts both ways. A detected watermark on your deliverable isn't evidence you didn't do the work, and clean-scanning content isn't evidence a human made it. Any client, employer, or platform treating these signals as proof of authorship is misusing them, and that misuse is where the first real disputes will come from.
One fear worth retiring: the watermark carries no identifying information about users, organizations, or conversations. It isn't a tracking mechanism pointed at you. Anthropic also reports negligible effect on model speed and no change to what it costs to use Claude.
How marketers actually use Claude, and where the mark lands
This stays abstract until you map it onto real work. Six patterns cover most marketing operations, and the change lands differently on each.
In editing and polish, a human writes and Claude tightens — 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.
With first drafts from a human brief, you supply the outline, positioning, key points, and sources, Claude produces a draft, and you rewrite and fact-check. Here the mark is arguably fair, since AI generated the prose.
Volume variations — twenty ad headlines, ten subject lines, five hooks — are close to pure generation, and among the highest-value AI uses in marketing. Few people would claim otherwise.
Repurposing turns a webinar transcript into a blog post, a blog post into a carousel, a case study into an email sequence. The content is yours; AI is doing format conversion. Research and synthesis — summarizing competitor pages, pulling themes out of customer interviews — often never gets published at all; it informs a human who writes something original. Translation takes your approved copy into another language, with a fully human source and a model-processed output.
In four of those six, the human supplied the substance and AI handled mechanics. Only first drafts and volume variations involve AI producing the ideas. A provenance mark looks identical across all of them.
Which is 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?
For files, yes. C2PA is an open standard and the Content Authenticity Initiative runs a free inspection tool at contentcredentials.org. Upload a file and see any signed provenance data attached. Running a few of your recent deliverables through it tells you what you're actually shipping.
For text, not yet, though that's changing. Anthropic has said it will offer a watermark detection API without providing implementation details or a date. Until it lands, 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 different things. Style detectors guess from word patterns and are famously unreliable, flagging human writing as AI and missing AI writing constantly. Anthropic's watermark is a deliberate embedded signal that external detectors have no key to read. If a tool can't tell you which of the two it's performing, don't trust its answer.
What to do
Find out what your stack runs on. Plenty of marketing platforms use AI under the hood without disclosing which model. Whether your output carries a mark depends on that, and it will 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, and it's better to close that on your own terms.
Revisit any "100% human-written" claim in your proposals and decide what you actually mean by it — human-directed, human-edited, or no AI at any step. Write down the honest version now. Vague authorship claims age badly in a world with provenance metadata.
Don't rebuild your workflow over this. 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.
The bigger picture
Anthropic is among the first major labs to implement the EU framework at this scale, and the direction is set. Other providers signed the same code and are building their own marking systems. C2PA adoption keeps spreading. Provenance signals are becoming infrastructure the way HTTPS went from novelty to default.
For content teams that's clarifying more than threatening. The period of ambiguity about AI involvement is closing, and the teams least disturbed by that are the ones whose value never rested on the ambiguity — where the strategy, the judgment, and the quality are the product. If that describes your operation, this asks nothing of you beyond being accurate about your own process.
