Claude Watermarking: The Act That Can't Keep Up
- David Turner
- Aug 17
- 3 min read
Updated: Aug 22
Under the EU AI Act (2024), AI companies are obliged to mark text generated by their LLMs as of August 2nd 2026. This means that companies who signed the EU Code of Practice on Transparency (including Anthropic, Open AI and Google) are introducing ways to do so. Failure to comply with text-marking could result in fines of up to €15 million, or 3% of worldwide annual turnover.
Much coverage has been given to Claude’s announced rollout of watermarking. Yet in reality there’s actually very little to say.
How Claude Watermarking Works
Technically it’s a relatively simple concept, executed via the usual array of abstract mathematics and software inherent to matters of cryptographic authentication. A secret key is used by Anthropic to create a subtle bias in the choice of words it selects in the output. This bias factor influences which word the LLM will select next from its list of statistical possibilities. This in turn creates a pattern in the text generated, which is the watermark in question. If you have the secret key, then you can identify which bodies of text contain a watermark. And in the interests of privacy, the watermark can’t be used to trace the chat, person or organisation who prompted the text.
The concept itself is based on a Google DeepMind paper which describes a technique called ‘SynthID-Text’ to watermark LLM generated text.
The final thing to say with respect to technicalities, is that it’s not infallible.
Why Watermarking Fails in Practice
If the text output is a response to a prompt containing a logical/factual question, then there’s obviously little scope for any sort of bias. Maths, code, historical fact, all fall into this category.
It’s also easy to remove the watermark. Not by a light edit of the text - the pattern will persist. It needs a heavy rewrite to erase the watermark. If you try that manually, then there’s not much point to the exercise - you may as well have written it yourself anyway. But paste it into another GenAI with a prompt to rewrite, and the watermark will be removed in seconds.
Which leads us to the real question - what’s the point?
The Real Problem: Watermarking Can't Solve What Matters
The internet is full of blog posts, articles, thought leadership pieces, sales pitches, and endless social media posts. And hardly any of them get read (data estimates from SEO giant Ahrefs indicates that around 30-40% of the web is made up of low-effort blog posts, less than 2% of which receive organic traffic). Even if they do get read, it’s a skim read at best. It’s largely an irrelevance to the audience whether those pieces are AI generated or not. They’re just ‘content’ - filler. The building aggregate of the internet. The junk food that LLMs devour as they crawl the vast featureless plains of internet wilderness; simple carbohydrates for conversational training.
Google ‘Claude Watermarking’ and you’ll find search results full of AI-generated articles posted by AI experts, replete with AI generated SEO-friendly clickbait titles, writing about the subject of AI authenticity and adorned with AI-slop images. Pythonesque.
That leaves only professional and academic work as the use case. If a novelist or academic authors an essay, book or whitepaper using AI, then can’t be bothered to paraphrase it to remove the watermark, then they’re hopefully in a lazy minority.
Ultimately what we’re looking at is regulation that can’t keep up with the frenetic pace of the technology it’s trying to regulate. Written and signed-off with good intentions no doubt, but written in another era in terms of AI. Anthropic and the other vendors simply have to implement something to avoid being fined.
Do you need to worry about AI watermarking? Probably not. People know when they see AI slop, and equally they know when they’re reading a badly written article by a human, and likely don’t read either anyway. In the real world, nothing really changes.
David Turner is the founder of Kói, an independent technology consultancy advising investors, founders, and boards on AI strategy and regulatory readiness.
You can reach him at: enquiries@dkoi.design
© Kói Holdings Ltd 2026. All Rights Reserved.
