Every time something shifts in AI, there is a freak-out online. This week was no different.
Anthropic announced it is embedding an invisible, machine-readable mark in the text Claude generates. Within hours the reaction had sorted itself into three piles.
This is Big Brother.
This is going to bury us in Google Search.
Do I now have to disclose it every time I use AI on a web page or a collateral piece?
I went and read the actual documentation. All three are wrong, though the third one is wrong in a way worth understanding.
So here is the plain version.
On Tuesday, Anthropic signed the EU AI Act's Article 50(2) Code of Practice on Transparency. Every Claude model released on or after August 2 now carries the mark.
The mark is invisible to you. It survives copy and paste. It may survive some editing. Image files that Claude generates carry signed provenance data inside the file.
This is not EU-only. It applies worldwide, across the API, Claude.ai, Claude Code, Cowork, Claude Tag, and Claude running on AWS, Google Cloud, and Microsoft Foundry. Anthropic published how the marking works.
Detection tools for third parties have been announced. They have not shipped. Nobody can scan your website for this today.
Ask Claude to proofread a memo you wrote yourself, and what comes back is marked. Same if you paste your own draft and ask it to tighten the wording. The mark says a Claude model touched the text. It says nothing about who did the thinking.
Short passages often carry no detectable mark. Heavy editing breaks it. Paraphrasing breaks it. Converting the file can break it. Anything written on an older model was never marked at all.
No prompt. No client data. No account, no name, no company. It is not a tracker. It says one thing: a Claude model touched this.
It did not before and this does not change it. Google's announced provenance checks in Search cover images, video, and audio. Not your blog post.
Every time this subject comes up, someone asks whether AI-assisted content will hurt their search rankings. Short answer: no, and the watermark does not change that.
Google's own guidance has been consistent for three years. It rewards quality, not authorship method. It does not ask how content was made.
What Google does penalize is scaled content abuse — pumping out volumes of low-value pages to game rankings. That was already against the rules when people did it by hand. AI just made it cheaper to do badly.
The watermark is not a ranking signal. No search engine has announced using it as one. It is a provenance mark, readable by detection tools that have not shipped yet.
Being cited by AI answer engines works the same way, only more so. What earns a citation from ChatGPT, Perplexity, or an AI Overview is what earned trust before: real expertise, specific claims, clear structure, sources a machine can verify. Thin content does not get cited. It gets skipped.
Here is the honest read for a business owner. AI-assisted drafting is fine. AI-generated volume is the risk — and the risk was never a watermark. It was publishing things nobody needed.
This is the part I want you to sit with.
California's AI Transparency Act became operative on August 2, 2026 — the same day as the EU rule. Not a coincidence. California originally set January 1, then amended the law to line up with Brussels.
It requires large AI providers to embed provenance data in generated images, video, and audio, and to publish a free public detection tool. Penalties run $5,000 per violation, and each day counts as a separate violation.
Two honest caveats. California's law does not cover text. And today it lands on companies with more than a million monthly users — your vendors, not you.
But look at the schedule. In 2027, large online platforms must start detecting and displaying provenance data. In 2028, camera and recorder manufacturers must support it by default. And a pending amendment, SB 1000, would remove the million-user threshold entirely.
Anthropic is first to ship text marking. Google, Meta, Microsoft, and OpenAI have all signed the same EU code of practice. The direction is not in dispute.
Nothing changed about whether using AI is acceptable. It is.
What changed is that hiding it stopped being viable.
Transparency infrastructure was always coming. It arrived on a Tuesday, and most people misread it as an accusation.
If you already have a written AI policy, you woke up Wednesday with an advantage. Not because the policy blocks anything. Because you already decided what you do, and you can say so.
This is the question I keep getting, so let me answer it precisely.
You draft a proposal with AI help. You edit it. You send it to a client. Does something have to go in the footer saying AI was involved?
No. For almost everything you send a customer, there is no labeling requirement.
Read Article 50(4) closely. The text disclosure duty applies to text "published with the purpose of informing the public on matters of public interest."
That is a narrow category. A proposal is not that. A capabilities deck is not that. A service brochure, a quote, a client report — none of those are that. The obligation never attaches in the first place.
Nobody has to stamp "AI-collaborated" on a quote.
If you publish something meant to inform the public on a matter of public interest — an explainer on a regulation affecting your industry, commentary on a public issue — then the duty applies.
And there the exemption is exactly what you would want. It does not apply where the content "has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility."
So for that narrow slice: documented human review with a named person removes the labeling requirement. That is the trade. A process instead of a disclaimer.
Your chatbot. Under Article 50(1), if an AI system interacts directly with people, they must be told it is AI unless it is obvious. If you have a bot on your site talking to EU visitors, that one is yours today, not your vendor's.
Realistic synthetic media. Images, audio, or video that could pass as real people or real events get disclosed.
Proposal, quote, client report, sales deck, brochure — No. Not "informing the public on matters of public interest." The rule does not reach it.
Blog post or explainer on a public issue or regulation — Yes, unless a named person did a documented review and holds editorial responsibility.
Chatbot on your website — Yes. Tell people they are talking to AI.
Realistic AI images, audio, or video of people or events — Yes.
Internal docs, drafts, notes — No. Not published.
Applies to EU-facing work. Useful everywhere.
1. Find where AI touches client-facing work. Deliverables. Proposals. Published marketing. Training materials. Your website chatbot. You are not hunting for problems. You are drawing an honest map.
2. Write down which bucket each one falls in. Commercial document, public-interest publishing, or direct interaction. The rules differ, and most of what you produce lands in the first bucket, where nothing is required.
3. Name who reviews AI-assisted work and signs off. One person, named, on paper. It is what unlocks the exemption for anything in bucket two, and it is plain good practice for everything else.
Here is the part worth remembering. None of this stops you from disclosing anyway. More RFPs and procurement questionnaires are asking how vendors use AI. Having a written, considered answer beats improvising one under deadline.
Said plainly: any policy template is a starting point. Have your own counsel review it before you rely on it.
We built aipolicy.johnconstantine.ai for exactly this moment. Two inputs: your company website and your work email. It reads your site, works out your industry, and pulls in the regulations that actually apply to you — including the marking and disclosure rules above.
Free. Download as Word or PDF, with an employee acknowledgment form included.
Then hand it to your attorney. It is a strong starting template, not a substitute for counsel.
Already a client? Your policy generator now includes a section called "AI Content Marking, Watermarking, and Disclosure" covering the Article 50 obligations above. If your policy was generated before August 2026, it predates these rules — re-run it and compare, or get in touch and we will read your current policy and tell you what is missing. No charge either way.
One email every two weeks — practical AI for small business owners. Free, and the forwarding kind of good.
Subscribe free →P.S. — Under Article 50, what exempts your published work from mandatory disclosure is a human review with a named person responsible. Not a tool. Name the person this week.
P.P.S. — This issue was written with the help of AI and reviewed by a human. That human was me. Per the section above, it does not need a label. I am telling you anyway.