Somebody is going to forward you an article this week, and the subject line will be some version of “is this going to hurt our website?”
On 10 August, Anthropic updated a support page confirming that Claude now embeds an invisible, machine-readable watermark into the text it generates. Within about a day, that turned into “AI content is going to get penalised” and started making the rounds.
So we did what we’d tell any client to do. We went and read the source material — Anthropic’s actual documentation, and Google’s published position going back to 2023. Not the coverage. The pages themselves.
Here’s the short version: the panic is misplaced, and the real problem is one nobody is talking about.
A watermark tells you a machine touched the text. It tells you nothing about whether the text is true.
That distinction matters more here than almost anywhere else, and we’ll explain why.
What did Anthropic actually announce?
Anthropic signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. The watermarking rollout is how it’s meeting that commitment.
Two mechanisms. Generated text carries an imperceptible watermark woven into the words themselves — invisible to you, unchanged in meaning, and because it lives in the text it travels when the text is copied and may survive some editing. Generated files like .png, .jpg and .svg get signed provenance metadata under the C2PA standard.
Three details worth knowing:
- It’s worldwide. The obligation is European; Anthropic is applying it everywhere Claude is offered. Trinidad included.
- Coverage isn’t complete. Models launched on or after 2 August 2026 mark at launch. Older models are still being brought in.
- Detection doesn’t exist yet. Anthropic says it will support third-party detection and publish technical documentation. That documentation is described as forthcoming. Today, nobody outside Anthropic can reliably check.
Will Google penalise your site for using AI?
No. And the fastest way to settle this is a fact most of the coverage skipped.
Google has been watermarking its own AI-generated text since 2024, through DeepMind’s SynthID. It signed the same EU transparency code Anthropic did, three weeks earlier. So Google has had this exact capability in-house for two years — applied to its own Gemini output — and in those two years has never made a text watermark a ranking signal, never mentioned watermarks in its content guidance, and never suggested marked text ranks differently.
Google’s published position since February 2023 is that appropriate use of AI or automation is not against its guidelines, and that its focus is content quality rather than how content was produced. What it does penalise is scaled content abuse — generating many pages primarily to manipulate rankings rather than help people. In 2024 Google restated that this applies “no matter whether content is produced through automation, human efforts, or some combination.”
Read that again, because it’s the part people miss. The policy is deliberately method-agnostic. Twelve human writers producing thin, templated pages break it. One carefully researched, AI-assisted article that genuinely answers a question does not.
Provenance is a disclosure question. Ranking is a quality question. This week the two got collapsed into one.
What does a watermark actually prove?
Considerably less than the headlines implied — and Anthropic’s own documentation is admirably precise about it.
When a mark is detected, it indicates the content may have been processed by Claude. Not written by. Not generated by. May have been processed by.
Anthropic spells out why. People use these tools to proofread, translate, summarise and convert files. The output can carry a mark even when the underlying ideas, text and data came from somewhere else entirely.
Think about what that means locally. A columnist writes 1,200 words from her own reporting, her own sources, her own argument — then runs it through for a grammar pass before filing. Marked. A translator working from a human-written original. Marked. Someone who pasted in a finished draft only to ask for a tighter headline. Marked.
And it fails in the other direction too. Anthropic states plainly that the absence of a mark does not mean content wasn’t AI-generated. Heavy editing strips it. Short passages don’t carry enough signal. Older models never marked anything. File metadata disappears through format conversion, re-saving, or a screenshot.
So anyone who tells you a watermark proves authorship is describing something the technology does not do. You can say that with confidence, and point at Anthropic’s own page when you do.
The slop is real. That’s not the argument.
Let’s be honest about what’s actually bothering people, because it isn’t rankings.
You’ve seen it. The fete flyer with the brush font and the food that doesn’t exist. The fundraiser medallion designed by nobody. The restaurant ad where the doubles look subtly, unplaceably wrong. Captions with that uncanny sameness — technically fine, saying nothing, all in the same voice that belongs to no one. Artwork without an artist.
That frustration is legitimate. Something genuine is being cheapened, and people can feel it before they can name it.
But here’s the thing about slop: you can see it. It announces itself. Your eye catches the too-smooth gradient, the sentence that means nothing, the sixth “quietly” in eight hundred words. You discount it and move on. It’s tacky, it’s lazy, and it’s visible.
The watermark will catch some of that. Fine.
It will not catch the thing that actually costs you.
What is the real risk of AI for my business?
Nothing marks a claim as false.
There is no watermark on a wrong price. No provenance metadata on the wrong opening hours. No signature on a confident, fluent, entirely fabricated sentence about your business, delivered to a customer who asked an assistant where to go tonight and never thought to check.
That’s the exposure. Not that AI wrote something. That AI said something — about you — that was never true.
And we have a first-hand example, from our own site.
We published something false. Here’s how it happened.
Earlier this year we published an article about Trinidad and Tobago’s AI readiness. In it, this sentence: “around 69% of Caribbean employers say their teams lack the AI skills they need.”
It read well. It supported the argument. It was wrong.
The real statistic comes from DeVry University’s Bridging the Gap (2025) — a survey of workers and hiring decision-makers in the United States, subtitled “Overcoming a Silent Standoff in America’s Talent Economy.” The finding is that 69% of US employers lack confidence their workers’ skills will translate over the next five years. Different country. Different question. Nothing to do with the Caribbean.
Now the part that should worry you more than the error itself.
When we went back to check the figure, an AI-generated summary told us — in clean, confident, unhesitating prose — that DeVry’s report found “69% of Caribbean employers believe their teams currently lack the AI skills needed.” It had quietly swapped the region and altered the finding.
That summary is almost certainly how the mistake got into our article in the first place. And it nearly caused us to put the false figure back. We caught it only because we opened the primary document instead of trusting the summary.
No watermark would have helped. The sentence was fluent, plausible, well-formatted and completely wrong. Provenance marking would have told us a machine was involved. It would not have told us the number was borrowed from another continent.
We removed the claim, replaced it with a verified statistic, and added a rule to how we work: a statistic travels with its citation, and a summary of a source is never the source.
Why are Caribbean businesses most at risk?
Because when regional data is thin, the vacuum doesn’t stay empty. It fills with numbers from somewhere else.
This isn’t a hunch. ECLAC’s own regional study — Artificial Intelligence Readiness in the Caribbean (2025) — opens by conceding that “data and research on the impacts of AI and AI readiness are limited” and that the study “is therefore not comprehensive in its assessment.” It states that “no agreed measures exist for the assessment of AI skills.” And it reports that individual country data on digital skills is available for exactly three Caribbean countries.
The World Bank puts it more bluntly: Caribbean small island states are among the most data-deprived countries in the world, with statistical capacity below that of low-income countries.
So when someone asks an AI assistant about Caribbean business, AI readiness, or your particular company, it answers from whatever it can find. Where there’s nothing, it reaches for the nearest plausible thing — usually American, usually stated with the same confidence as a fact.
Every year the region goes unmeasured, the volume of confident misinformation about it grows.
Do we use AI ourselves?
Yes — and it would be dishonest to write this without saying so plainly.
We use AI daily: research, outlining, drafting, editing. This article was written that way. Which is exactly why we’ll tell you what that does and doesn’t mean.
Every factual claim here was checked against a primary source: Anthropic’s own support documentation, Google’s published guidance, ECLAC’s report, DeVry’s release. Not blog coverage. Not summaries. The pages themselves. Claims that couldn’t be traced to a primary source were cut rather than softened.
A model helped us write it faster. It could not have told you about the 69% error, because that came from our own work and our own embarrassment.
That’s where the line sits. AI-assisted, human-directed, human-accountable. The tooling changed how quickly a draft appears. It did not change who is responsible for whether it’s true.
Google’s own guidance, incidentally, recommends disclosure where a reader might reasonably wonder how something was made — and advises against giving AI an author byline. That seems right to us.
What should my business actually do about it?
Stop worrying about whether AI wrote it. Start worrying about whether AI can read you, and whether what it says about you is correct.
Practically:
- Find out what AI currently says about you. Most businesses have never checked. The answer is often out of date, incomplete, or describing a business they stopped being years ago.
- Make the facts easy to verify. Consistent name, address, phone and hours across your site, your Google Business Profile and every listing. Contradictions are what make an assistant hedge — or guess.
- Put your knowledge somewhere machines can read it. Answers to real questions, in public, in plain language.
- Check your own claims. If you publish a statistic, publish where it came from. We learned this the hard way and we’re telling you so you don’t have to.
The watermark debate will be over in a fortnight. The question of whether AI can accurately describe your business will matter for years.
Where to start
Find out what today’s assistants actually say about you. Our free AI Visibility Check asks them, shows you the answers, and tells you where you’re invisible or misdescribed — no charge, no obligation.
Run your free AI Visibility Check →
Frequently asked questions
Does Google penalise AI-generated content?
No. Google’s published position since February 2023 is that appropriate use of AI or automation is not against its guidelines, and that it judges content on quality rather than how it was produced. What Google penalises is scaled content abuse — generating many pages primarily to manipulate rankings while providing little value — and that policy applies equally to content made by humans, by automation, or by a combination.
Can Google detect Claude’s watermark and use it against me?
There is no evidence it does. Production method does not appear as a ranking input anywhere in Google’s published documentation, and Anthropic’s detection tooling has not shipped. Notably, Google has watermarked its own Gemini text output since 2024 and has never tied watermarks to rankings in that time.
Does a watermark prove that AI wrote something?
No. Anthropic’s documentation states that a detected mark indicates content may have been processed by Claude, and lists proofreading, translating, summarising and file conversion as cases where a mark appears even though the ideas and words came from elsewhere. It shows a tool was involved at some stage. It says nothing about who wrote the underlying work — or whether it is accurate.
Should my business stop using AI?
Google’s guidelines do not require it, and avoiding it entirely just costs you time. The workable approach is AI-assisted drafting with human direction, independent verification of every factual claim, and a person accountable for the result before anything is published.
What is the real risk of AI for Caribbean businesses?
Not being penalised — being described wrongly. AI assistants answer questions about local businesses using whatever information they can find, and Caribbean data is scarce enough that they often fill gaps with figures and assumptions from elsewhere. If what AI says about your business is inaccurate or out of date, that is the problem worth fixing.
Sources
- Anthropic — How Claude marks AI-generated content
- Google Search Central — Google Search and AI-generated content (February 2023)
- Google Search Central — March 2024 core update and new spam policies
- Google DeepMind — SynthID
- European Commission — Code of Practice on Transparency of AI-Generated Content
- ECLAC / United Nations — Artificial Intelligence Readiness in the Caribbean (2025)
- World Bank — Closing the Caribbean Data Gap
- DeVry University — Bridging the Gap (2025)