On Tuesday, 11 August, Anthropic confirmed a change this column has been expecting for a while: text generated by its newest Claude models now carries an invisible watermark, woven into the words themselves, everywhere Claude is used, worldwide. Supported files such as images get a signed provenance label on top, under the industry C2PA standard. The trigger is European law — the EU AI Act’s transparency rules became enforceable for newly launched AI systems on 2 August — but rather than run a separate European version, Anthropic is applying the marking globally.

Nor is this a Claude quirk. OpenAI, Google, Microsoft, and Meta have signed the same European code of practice, and platforms have begun flagging AI content on their own initiative. The direction is settled: AI-generated content is becoming identifiable. Any business that has built AI into its weekly routine, as this column has been encouraging for ten weeks, should understand what that means before the detection tools land in other people’s hands.

New Claude models watermark their text at the model level, wherever Claude is used — with
signed provenance labels on files beyond text. (Branded graphic by PGH Consulting, LLC)

What actually changed

Claude models launched on or after 2 August 2026 mark their output from day one; Anthropic says marking for its older models is being worked on during a transition period the law allows. The watermark is applied at the model level, which means it is present whether you use Claude in the browser, through the API, in Claude Code, in the Cowork desktop tool this column covered in Week 5, or through the big cloud platforms.

The mark itself is not a visible stamp. It is a statistical pattern in the choice of words, imperceptible to a reader, that specialised software can test for. Anthropic says it does not change the meaning, quality, or readability of the text. Because the pattern lives in the text itself, it travels with copy and paste, and it can survive a degree of editing. How much editing removes it, Anthropic has not said. Detection is the part still in progress: the company has committed to helping users and third parties check content for Claude’s marks, with technical documentation promised. So “everyone can tell” is where this is going rather than where it is this morning — but the text you generate today carries the mark that tomorrow’s tools will read.

Not a visible stamp: a statistical pattern in the words themselves that travels with copy-paste
and may survive light editing, with detection tools to come. (Branded graphic by PGH Consulting, LLC)

What a detected mark proves — and what it does not

Anthropic is unusually plain about the limits, and they matter more than the headline. A detected watermark proves that text passed through Claude. It does not prove Claude did the thinking. Paste your own management report in for proofreading and the polished version can carry the mark, even though every idea and every figure is yours. Ask for a translation of a letter you wrote and the translation is marked. The mark records processing, not authorship.

A detected mark proves text passed through Claude, not that Claude did the thinking — and an
unmarked document proves nothing at all. (Branded graphic by PGH Consulting, LLC)

The absence of a mark proves even less. Text from an older model, text that has been heavily edited or paraphrased, a passage too short to hold a reliable signal, an image passed through a screenshot — all can read as unmarked. So the watermark is a piece of context, not a verdict, and anyone who waves a detector at a document and announces fraud has misunderstood the tool. That cuts both ways: it protects you from unfair accusations, and it should temper how you judge what others send you.

What changes for a business using AI properly

For the workflow this column has taught since Week 1, almost nothing. The figures still come from your books, because AI reads figures and never generates them. The verification still happens, the judgement is still yours, and the signature at the bottom is still yours. None of that is touched by a watermark.

What disappears is deniability. From now on, the working assumption for every proposal, report, and letter is that AI involvement can eventually be identified. If that is uncomfortable for a particular document, the fix is not hunting for an unmarked tool; it is repairing the claim you were making. A business that markets “hand-crafted, no AI” while drafting with a chatbot now has a problem with a timestamp on it. A business that says “we use AI carefully and check everything” has nothing to walk back.

The staff dimension deserves a moment. Last week’s column argued that a one-page AI policy beats a ban, because banned tools migrate to personal phones. The watermark strengthens that argument considerably: work produced against the rules on a personal account is no longer invisible, and no manager wants to discover policy violations by scanning documents after the fact. Give people approved accounts and clear rules, and the watermark becomes a non-event.

The other side of the desk

In time, the checking runs both ways. The tender response that reads suspiciously smooth, the glowing reference letter, the supplier proposal — once detection tools are public, you will be able to test what you receive. Use that power with the limits above in mind: a mark may mean the sender used AI to translate or tidy something they wrote themselves, and a clean result may mean nothing at all. This is also different from the “AI detector” websites that guess from writing style and are wrong often enough to have hurt innocent people: a designed-in watermark, checked mathematically, is far more reliable — but only for what it actually claims to show.

Fix the claim, not the wording: a one-line disclosure, a transparency rule in the office AI policy,
and the same verification habits as always. (Branded graphic by PGH Consulting, LLC)

Disclosure was already the right habit

This column’s carousel images have carried an “AI-generated” credit line since the first week, and the sky has not fallen. Disclosure is a sentence, not a confession: “prepared with AI assistance and reviewed by us” costs nothing when trust in the underlying work is deserved. Where the reader reasonably assumes a human wrote every word — published writing, academic work, anything contractual that says so — say so plainly. Everywhere else, the standard being set by regulators, and now by the technology itself, is simply that you should not claim otherwise. The businesses that have been transparent all along get to change nothing this week, which is its own small reward for doing it right — and the rest have been handed a very good reason to start.

What to try this week
1. Read Anthropic’s support note “How Claude marks AI-generated content” — it is short, plainly written, and the source of record for what is and is not marked.
2. List the documents that leave your business with AI’s help — emails, proposals, reports, social copy — and ask of each: would identification as AI-assisted embarrass us? Fix the claim, not the wording.
3. Add a one-line disclosure where readers reasonably assume fully human work: “Prepared with AI assistance and reviewed by [name].”
4. Update the one-page AI policy from last week’s column with a transparency rule: nobody describes AI-assisted work as AI-free, inside or outside the business.
5. Carry on exactly as the series has taught: your figures, your verification, your judgement. The watermark changes who can tell AI was involved — it does not change who is responsible for the work.
Detection tools are not yet in the public’s hands, a mark proves processing rather than authorship, and an unmarked document proves nothing at all — so do not accuse, and do not panic. Disclose, verify, and sign your work like you mean it.

Peta-Gaye Hardy is the founder of PGH Consulting, LLC, where she helps finance and operations teams adopt AI in practical, low-risk ways. She writes the weekly AI in Finance & Business column and is based between Jamaica and the United States. Learn more at www.pghconsultinggroup.com. Follow on Instagram and YouTube @pghconsultinggroup, and connect on LinkedIn at linkedin.com/in/peta-gaye-hardy.

Disclosures: This article is informational and does not constitute investment, tax, legal, or accounting advice. Details of Anthropic’s content-marking approach are as published by the company at the time of writing (12 August 2026) and are subject to change as its technical documentation is released; readers should confirm the current position on the vendor’s own pages. AI tools can produce errors, and every figure, clause, or claim they produce should be verified against a source before being shared or acted upon. The author has no commercial relationship with Anthropic, OpenAI, Google, Microsoft, Meta, or any product mentioned and was not compensated by them. The examples described are illustrative and do not depict any real business.