Big news from the robots: the text they write now comes with a watermark.
Two months ago, on 2nd August 2026, Article 50 of the EU AI Act kicked in, requiring AI companies to mark their outputs in the form of a watermark. Anthropic (the company behind Claude) promptly announced that its new models would embed this invisible watermark in every piece of text they generate – not just in the EU, but everywhere, because they haven’t worked out how to switch it off by region. Google’s Gemini has already been doing this since 2024, and OpenAI, Meta and Microsoft have all signed up to the same EU code of practice and are in the process of implementing it.
The internet lost its collective mind and almost every “hot take” you read about the topic falls into one of two camps:
Camp One: AI evangelists declaring that the watermark is completely meaningless and easily stripped. Nothing to see here, please continue generating 400 boring blog posts a week.
Camp Two: copywriters doing a smug little jig and shouting “I TOLD YOU SO. Hire a human. Preferably me.”
I’m an award-winning copywriter – the best-rated in Lincolnshire – and I also openly use AI in parts of my workflow. So I’m about to do something neither camp will enjoy: tell you the truth.
What is the AI watermark?
The most important thing to understand is that it’s not a watermark in the traditional sense of the word. And to be honest I think this is where a lot of the confusion stems from.
It’s not hidden characters. It’s not a logo. It’s not something you can see, find, or delete.
The watermark is a statistical pattern baked into the word choices themselves. When the AI writes, a secret key subtly nudges it towards some equally-good words over others, creating a fingerprint that a detector holding the same key can spot later. Because the watermark IS the words, it travels with the content when you copy and paste it, and it even survives light editing.
Think of it this way. Imagine you’re a novelist and every time you have a choice between two words that would do equally well you roll a loaded dice to tell you which one to pick. The dice is rigged to favour certain words from a secret list – say, words whose first letter falls in a particular pattern. “Big” or “large”? The coin says “large”. “Begin” or “start”? The coin says “start”.
Every sentence looks normal and reads well because each choice is a completely reasonable choice that a human could easily have made. But if you know the secret list and you count across a thousand words, you will inevitably find the favoured words turning up far more often than ordinary chance would allow. That statistical fingerprint is the watermark.
We can spot it to some degree. The “it’s not X, it’s Y” sentence structure; the em dashes; odd words like “quietly”, “delve”, “unlock”; phrases like “here’s the thing”, etc. A bit of light editing will get rid of the human-detectable tells, but not the full watermark.
The watermark proves involvement, not authorship
One unfortunate truth with all this is that the watermark reveals to you that an AI has touched the text. But it does not tell you that an AI wrote it.
Anthropic says this itself: the watermark “cannot distinguish ‘Claude wrote this’ from ‘Claude heavily edited this'”. If you write a beautiful, first draft entirely by yourself and then ask an AI to proofread it, translate it, or tidy up your formatting then it’ll likely get watermarked. The content could be entirely made up of your ideas and your expertise, but it’ll get stamped with the same branding as fully machine-generated AI slop.
And the reverse is also true: no watermark does not mean no AI was used. Older AI models don’t watermark yet. Heavily rewritten AI text will lose the watermark. Short passages won’t carry enough of a signal for the mark to be detected. An absent watermark proves precisely nothing about human authorship.
So if someone does decide they’re going to feed your content through a detector and gets a result – either result – they’ve not really learned anything at all.
Can I check my content for a watermark?
Watermark detection is only available to eligible organisations (such as regulators, law enforcement, media, fact-checkers, and compliance-obligated enterprises) as required under EU law.
There may be plans for a future roll-out, but right now the AI companies are the only ones who hold the key to reading the watermark. So any “Watermark Detectors” you find online are either guessing from the writing style or scanning for hidden Unicode characters, which is a whole different kettle of fish that has nothing to do with watermarking at all.
The one exception is files. Images and graphics generated by AI now carry signed provenance metadata using an open standard called C2PA, and that can be verified today with free tools. But there’s currently no way to check text.
Can an AI watermark be removed?
Yes – but not in the way you’re probably thinking.
Light editing won’t remove it and the tools claiming to “strip” it by deleting invisible characters are, as we’ve established, not removing it at all. The only way to actually remove the watermark is to rewrite the text so extensively that pretty much every word gets replaced.
At which point it’s arguable whether the text can even be considered AI-generated any more. You’ve technically written it yourself by that point.
Which is why my advice isn’t “here’s how to launder your AI content”. It’s the opposite: if your content strategy depends on nobody ever finding out how the sausage was made, the problem isn’t the watermark. The problem is the sausage.
Will a watermark hurt my SEO?
Yes and no. The watermark itself doesn’t affect your Google rankings, but bland and generic content does. And most AI content is bland and generic.
Search engines reward helpful, original, expert content however it’s produced, and they penalise low-value content churned out at scale in an attempt to game the SERP rankings. The watermark changes neither of those things.
You will see a study doing the rounds claiming watermarked content ranks about five positions lower than un-watermarked content. Read the small print: the researchers themselves admit they didn’t control for quality. The human-written content in the study may simply have been better.
Could Google use watermark signals in future? Theoretically. But the thing that’s actually tanking AI content in search results is that most of it is thin, generic, and identical to the forty other articles generated from the same prompt.
Fix that and the watermark is irrelevant. Don’t fix it and the watermark is the least of your problems.
So what does this mean if you’re using AI for your marketing?
If you’re a business owner using AI to help with your content – and let’s be honest, many of you are, and that’s fine – here’s what has actually changed:
1. You should assume everything is traceable, because eventually it will be. The detection tools are coming. You should build your content process as if anyone could check it, because within a year or two, anyone will be able to.
2. Stop treating AI use as a dirty little secret. The watermark era makes “we never use AI” an increasingly risky claim and an increasingly pointless one. And to be honest, saying you won’t use AI is not the brag you think it is. What clients and customers actually care about is whether the thinking is yours. Decide what your AI policy is, write it down, and make sure you feel comfortable with it.
3. Watch your contracts. If you’ve promised a client (or been promised by an agency) “100% human-written content”, either mean it or renegotiate it. The watermark doesn’t punish AI use, it punishes lying about it. And if you lie now then you’ll get caught out further down the line when the detectors go live.
4. Don’t panic-buy detection or removal tools. As we’ve covered: the detectors can’t detect and the removers can’t remove. Save your money for something that works, like a decent editor or a copywriter (hellooooo there!).
My honest opinion
The watermark won’t eradicate AI content. It won’t suddenly affect your rankings, catch out cheats, or settle an argument about authorship. But that’s fine because it was never really designed to do any of those things. It’s a transparency measure, and a tentative one at that.
What it will do is drag a question into the open that most businesses have been dodging: what is your content actually for, and who is doing the thinking behind it?
If the answer is “an intelligent human with expertise” then nothing has changed for you.
If the answer makes you a bit hot under the collar, the watermark isn’t your problem. But I might be able to help with the thing that is.
Get in touch and let’s have a chat.