From 2 August 2026, Anthropic began marking Claude’s output so it can later be identified as machine-generated. If your business uses Claude to draft blog posts, product descriptions, newsletters or social captions, much like the workflows built into ChatGPT’s new small business agent, the fact that Claude watermarks its output is now something worth understanding, even if none of your customers will ever see it. Many owners are also exploring free AI training built specifically for small business owners. Anthropic confirmed the change in an announcement on 11 August 2026, and it applies everywhere Claude runs, with no way to switch it off. For a look at the risks AI tools can introduce if left unmanaged, see how an autonomous AI recently compromised a small business.
Here is the short version for a busy owner: this changes the provenance of your content, not its legality. For most UK small businesses publishing their own marketing copy, nothing you are doing suddenly becomes a problem. It’s still worth setting clear rules on AI use, as outlined in our practical guide to writing an AI usage policy. It is still worth having a written AI usage policy that sets out how AI-drafted content gets checked before publication. If you’re weighing up AI tools more broadly, see how the ChatGPT small business agent with free training compares. But it is a good moment to tidy up how you use AI, including tools like the ChatGPT small business agent with free in-person training, and to understand what the watermark is and is not.
What Anthropic has actually done
Two things now happen automatically when Claude produces content.
First, text output carries an invisible watermark woven into the text itself. It does not change the meaning or the readability of what Claude writes. It works by subtly biasing word choice, so it becomes detectable across a body of content rather than in a single sentence. Importantly, it survives being copied and pasted, so lifting the words into your website builder does not strip it out.
Second, files that Claude produces carry signed provenance metadata using a standard called C2PA (the Coalition for Content Provenance and Authenticity, an industry standard backed by the likes of Adobe and Microsoft). This is a tamper-evident record attached to a file that says where it came from.
The change is global, with no opt-out, and it covers the Claude apps, the API, Claude Code, Claude Cowork and Claude Tag, a reminder that AI tools can carry risks beyond content labelling, as shown when an autonomous AI hacked a small business. In other words, if the words came out of Claude, they are marked, much like the transparency prompts now built into ChatGPT’s small business tools.
Why now? The EU AI Act is the real driver
This is not Anthropic being cautious for its own sake, though it sits alongside other AI risk stories such as how an autonomous AI hacked a small business. The driver is regulation. The EU AI Act’s Article 50 sets a transparency duty, and its Transparency Code took effect on 2 August 2026. That duty requires AI-generated or AI-edited content to be marked so that other systems can identify it. Watermarking is how a provider like Anthropic meets that obligation on the content its model produces.
The key word there is provider. The duty in question falls on the companies that build and supply the AI models, not on every small business that happens to use one. More on what that means for you below.
Can a customer see the watermark?
No. In normal reading, the text watermark is invisible. A customer scanning your services page or reading your newsletter cannot see it, cannot feel it, and would need specialised detection tooling to find it. It does not add a badge, a disclaimer or any visible mark to your copy.
What it does mean is that, in principle, someone running the right detection software across enough of your published text could infer that Claude was involved. That is a provenance signal, sitting quietly in the background. It is not a warning label your audience will trip over.
Does Google penalise AI content because of the watermark?
This is the question most owners will worry about, so let us be precise.
Google’s stated position is that it rewards helpful, original, people-first content however it is produced, and that it acts against unhelpful, mass-produced content made primarily to game search rankings. In Google’s own words, using AI to help create content is not against its guidelines; churning out low-value pages at scale is.
A watermark is a provenance signal, not a ranking penalty. There is no basis for claiming that Google reads Anthropic’s text watermark, or that it ranks pages up or down based on it. If your content is genuinely useful, well edited and accurate, the presence of a watermark does not change how it should perform. If your content is thin, repetitive filler, it was already at risk, watermark or not.
The practical takeaway is unchanged from the advice we gave when we looked at how AI adoption lifted profitability at top UK accounting firms: the firms that win use AI to do better work faster, not to flood the internet with words nobody asked for.
Do I have to disclose that I used AI?
Here is where you need to separate two different things.
The EU AI Act’s transparency duty is primarily aimed at AI providers and at specific high-risk or public-facing uses, such as deepfakes and content on matters of public interest. It requires that AI-generated material be marked in a machine-readable way, which is exactly what Anthropic has now done on its side. It does not impose a blanket legal requirement that every UK small business add “written with AI” to the bottom of every blog post about its plumbing services.
Whether any disclosure obligation reaches your particular business depends on circumstances: what you publish, who you publish it to, whether you serve EU customers, and whether the content touches areas the Act treats as sensitive. If you are marketing to consumers in the EU, or producing anything that could be mistaken for genuine reporting or a real person’s likeness, take proper advice rather than relying on a general article. For the everyday case of a UK firm writing its own marketing copy with AI help, this remains a provenance change, not a new legal trap.
There is also a broader consumer-protection principle worth remembering in the UK. Under advertising and consumer law overseen by the Advertising Standards Authority and the Competition and Markets Authority, you must not mislead. That principle already applied before watermarking existed. An AI-written testimonial from a customer who does not exist was always a problem. AI-assisted copy describing a real service honestly is not.
How Claude compares with the other AI tools you might use
Claude is not the only model marking its output, and the approaches differ. If you use more than one tool, it helps to know where each stands. The table below reflects what each provider is generally known to do; specifics evolve, so treat it as a starting shortlist rather than the last word.
| Tool | Provider | Approach to marking output |
|---|---|---|
| Claude | Anthropic | Invisible text watermark plus C2PA metadata on files, global, no opt-out from August 2026 |
| Gemini and Imagen | Google DeepMind | SynthID watermarking, best known for images, audio and video, extended to text |
| ChatGPT and DALL-E | OpenAI | C2PA provenance metadata on generated images; a member of the C2PA coalition |
| Copilot | Microsoft | Content credentials via C2PA on generated images, part of the same provenance push |
Separately, a small industry of AI detection tools has grown up on the other side of the fence. Tools such as Originality.ai, GPTZero and Copyleaks claim to spot machine-generated text. These are not watermark readers as such; they make probabilistic guesses, and their accuracy is contested. If a client or publisher runs your work through one, the watermark and the detector are doing different jobs. Neither is a reason to panic, but both are reasons to make sure a human stands behind what you publish.
Can I just remove the watermark?
Light editing probably will not remove a text watermark, because the signal is spread across word choices rather than sitting in one phrase. A full rewrite that replaces every word will remove it, but at that point you have genuinely rewritten the text yourself, and it is arguably no longer AI-produced content anyway.
The honest conclusion is that trying to scrub the mark out is the wrong goal. If your process already involves editing AI drafts into something you are happy to put your name to, you have nothing to hide and no reason to fight the watermark. If your process is publishing raw output at volume, the watermark is the least of your worries.
Frequently asked questions
Will my website visitors notice anything different?
No. The watermark is invisible in normal reading and does not alter the meaning or flow of the text. Visitors experience your content exactly as they did before.
Does this mean Google will now know my blog was written by Claude?
There is no evidence that Google reads Anthropic’s watermark or ranks content on it. Google judges content on whether it is helpful and original. Treat the watermark as a provenance record, not a search signal.
I run a UK-only business. Does the EU AI Act apply to me?
It can, if you serve customers in the EU or your content reaches EU users, and the duties are heavier for sensitive or public-facing content. For a UK firm writing its own honest marketing copy for a UK audience, the practical impact is minimal, but take advice if you are unsure or if you operate across borders.
Should I add a note saying my content is AI-assisted?
You are not generally required to, but you should disclose wherever a reader would reasonably expect it, for example if you publish AI-generated imagery of a person or content that could be mistaken for independent reporting. Honesty is the safe default.
Does the watermark apply to Claude Code and other Claude products?
Yes. Anthropic has applied it globally across the Claude apps, the API, Claude Code, Claude Cowork and Claude Tag, with no opt-out.
What to do next
- Edit, do not just publish. Treat every AI draft as a first draft. The businesses getting the most from AI, as the Dell study on UK SME AI front-runners found, use it to work faster, then apply real human judgement before anything goes live.
- Keep a named human accountable for accuracy. Decide who signs off published copy and who owns the facts in it. A watermark does not make claims true; you do.
- Disclose where a reader would expect it. Add a simple note on AI-generated imagery or anything that could be mistaken for independent content, and keep your marketing claims honest under existing UK advertising rules.
- Review your AI workflow. If you are experimenting with tools like the new ChatGPT small business agent alongside Claude, write down how content moves from draft to publication and where a person checks it. That single document protects you far more than worrying about a watermark ever will.
