Your AI Content Is Now Signed. Plan Accordingly.

Anthropic published a support article this week describing how Claude marks the content it generates. It is written in the flat register of a compliance notice, which is probably why almost no one in marketing has read it. They should.

Two things are now true. When a supported Claude model writes text, it weaves an imperceptible watermark directly into the text itself — Anthropic's word is imperceptible, not hidden, and the mark is applied at the model level, meaning it is present regardless of which Claude product or surface the text came from. And when Claude generates a file — a .svg, .png, or .jpg — it attaches signed provenance metadata following the C2PA open standard, the same content-credentials framework the rest of the industry has converged on.

The watermark travels. Anthropic says it will move with the text when it is copied and pasted elsewhere, and "may persist through some editing."

Read that last clause slowly if you run a marketing team.

This is a standard, not a vendor quirk

The reflexive read is that one AI company did an unusual thing. It isn't.

Anthropic signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, and the marking work is how that commitment gets implemented. The EU's Article 50 transparency obligations apply broadly to providers of generative systems. C2PA — the Coalition for Content Provenance and Authenticity — is an open standard already adopted across camera makers, publishing platforms, and other model providers. California's AI Transparency Act (SB 942, as amended by AB 853) became operative in August 2026, requiring covered providers to embed latent disclosures and offer free public detection tooling.

Three separate legal regimes, one direction of travel. Provenance is becoming infrastructure — a property of the file and the sentence, not a claim on the page.

So the strategic question changed shape. It is no longer can anyone tell? The engineering answer to that is trending toward yes. It is whether the label arrives from you, in your own framing, or from a detector in someone else's hands.

Who can actually check

This is the part most coverage gets wrong, and it matters for planning.

For files, Anthropic offers a free public checker for Claude-issued content credentials. Anyone can verify whether an image carries the mark and whether it was tampered with.

Text is different. Watermark detection is in private preview, and access is deliberately narrow: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, EU civil-society groups, and enterprises with their own compliance obligations under the Act. Anthropic says it plans to expand access over time.

So today, your competitor cannot paste your landing page into a box and get a verdict. A regulator can. A journalist can. A large enterprise running a compliance check on a vendor can. And "over time" is doing real work in that sentence — the planning assumption should be that detection gets broader, not narrower, and that content you publish today remains checkable years from now.

Content is durable. Detection capability is expanding. Those two facts compound in one direction.

The exposure is not what people think

The instinct in most marketing organizations right now is to treat this as a risk to be managed — quieter AI use, less of a paper trail, a hope that the mark degrades in the rewrite.

That is the wrong posture, and not primarily for ethical reasons. It is a bad bet on the facts — you would be building a content operation on the assumption that a capability curve reverses, which it has not once done.

The actual exposure isn't that AI was used. Nearly everyone is using it, and buyers assume as much. The exposure is the gap between what a company said about its process and what the provenance data shows. A firm that never mentioned AI and gets asked about a flagged asset by a procurement committee has a credibility problem. A firm that published its standard eighteen months earlier has a footnote.

Same underlying facts. Entirely different conversation. The difference is disclosure, and disclosure is free — but only in advance.

What we do, plainly

A piece like this is worthless from someone hedging, so: Commonwealth Creative uses AI throughout our work — research, drafting, analysis, production. We have said so publicly for as long as we have done it, and this changes nothing about how we operate, because there was never a version of our process built on the assumption that nobody would find out.

What we hold to is the part the tooling can't do. A human with domain expertise sets the direction, checks the facts, and puts their name on the result before it goes anywhere. Every factual claim traces to a source. Nothing publishes without a person accountable for it.

That is not a compliance posture. It is why the work is any good. A model that has never met your customers cannot know which of the true things is the one worth saying.

What to do this quarter

Four things, none of them expensive.

1. Write down your AI standard. One page. Where AI is used, where a human reviews, who signs off, what never ships without a person. Most firms have this in someone's head. Getting it on paper is a single afternoon and it is the artifact everything else depends on.

2. Publish it before anyone asks. A short, plain page on your site. Disclosure made voluntarily reads as confidence. The identical disclosure made under questioning reads as an admission. The content is the same; the timing is the whole asset.

3. Check what your vendors are doing. If an agency, freelancer, or contractor delivers content to you, their process is now your provenance record. Ask them the question directly. The ones doing it well will answer immediately.

4. Stop treating "made with AI" as the damaging fact. It isn't. The damaging fact is unreviewed, unsourced, unaccountable content — which was already damaging before anything was watermarked, just harder to prove.

The part worth sitting with

For three years the implicit deal in AI content was that the seams were invisible, and a great deal of low-quality work traveled on that assumption.

The seams are now legible — first to regulators and researchers, eventually to more. What that rewards is not the firms that hid it best. It is the firms whose process would survive the disclosure, and who said so while saying it was still a choice.

Provenance infrastructure is not a threat to companies doing the work properly. It is the first mechanism that has ever made the difference visible.

Sources: Anthropic — How Claude marks AI-generated content (accessed September 3, 2026); Claude Content Checker; California AI Transparency Act (SB 942, as amended by AB 853), operative August 2026.