Time to read: 10 minutes
A recent announcement from Anthropic says that text from Claude will now be watermarked. Here is everything you need to know about it: how the Claude watermarking system works, why you should care about it, and what to do if you’ve been writing content with Claude or working with an agency that does.
💡 Key Takeaway
If you’re publishing AI-generated text, watermarking makes it detectable – and light editing won’t hide it. The safest position is the one we’ve always taken at Sixth City: 75% human, 25% AI.
How Text Watermarking Works
The way LLMs build sentences is that they choose the next word based on probability. So in this example, we’ve got “the results of the study were quite…” And then what comes next? There are four candidate words, and it’s essentially a roll of the dice.
It shows the fourth candidate, which is “notable.” Now, the way the watermarking system works is that it’s going to basically choose a couple of these candidate words and mark them.
So I’ll go ahead and apply the secret key here to the candidates, and you can see that a couple of them were marked in green. This isn’t specifically how Claude’s system works, but it’s a great example to illustrate the mechanics of it.
So now if you have a long paragraph of text and it’s got a lot of these words that are highlighted in green, then you are watermarked, and you’re detected as using AI to generate the text.
Seeing It on a Real Passage
Let’s see how this works in an actual example. I’m going to make the text really long, because longer examples are easier to detect. And let’s go ahead and count this with the secret key on – you can see it’s lighting up like a Christmas tree.
We have a lot of these candidate words that are part of that secret key. It’s kind of like we’re caught red-handed by being highlighted in green, if you will. 70% of this text is green, and therefore we have passed a threshold. The watermarking system has shown that this was generated by AI.
Can You Edit Your Way Out of It?
One of the most common questions is, “Can’t we generate text with AI but then do some editing and some humanizing to evade the system?” With this system, it’s going to be actually really difficult.
The bottom line: every level of editing short of a full rewrite still leaves the text detectable.
| Level of Editing | Still Detectable as AI | Verdict |
|---|---|---|
| Fix the typos only | 96% | Plainly detectable |
| Light touch-up, some rewriting | 86% | Plainly detectable |
| Tighten sentences, rewrite more aggressively | 63% | Plainly detectable |
| Heavy edit / heavy rewrite | 25% | Borderline |
Source: watermark detection demo shown above.
And this is where we hit a wall. Because if you’re going to do a heavy edit or a heavy rewrite, what’s the point of using AI in the first place? Its purpose is efficiency and ease. You’re kind of losing that if you have to do the heavy rewriting, which is what makes this system really tricky.
⚠️ Warning
The only edit heavy enough to beat the watermark is a rewrite so extensive that it defeats the reason you reached for AI in the first place.
It’s also why at Sixth City Marketing, we’ve always had the philosophy that we go mostly human. We meet with our client experts, we use their quotes verbatim on their websites, and our writers go the human route as well. We don’t just plug a topic into Claude or ChatGPT and have it spit out the content for us to go and publish.
Why Is Anthropic Doing This?
From what I can tell, it’s mainly compliance. They say in this tweet that they’re implementing watermarking to comply with the EU AI Act. But I am a little concerned that the watermarking system might be used for other purposes as well.
I want to take you to this quote from Anthropic’s blog, answering the question, “How do I check if a piece of text was written by Claude?”
“We will soon be offering a watermark detection API.”
– Anthropic
Now, this is unknown. I don’t know if any search platforms, social media platforms, or whatever are going to tap into this API to help them detect what, if any, text was written by Claude. But that’s something that could be coming.
AI Labels Are Already Showing Up
I do see this sort of thing on other social platforms, like on Instagram. They don’t use a system like Claude’s watermarking, but they have their own thing going, and I’m seeing a lot more posts carry an AI info flag.
I know a lot of creators who are not happy about this, because they do use AI to help them build transcripts and things like that. But a lot of them are not using it too heavily, and it’s still getting flagged. Their audience is seeing this, and with the reputation that AI has in some spheres, you could lose trust.
Are we going to see something like this in the search results for Google or Bing or somewhere else? Are we going to see it on other social platforms? If you use Claude to write a LinkedIn post, are we going to see something? I’m not really sure, but it’s definitely something to be aware of.
Google Already Had a Hard Stance on AI Content
Here’s the thing, though. Let’s just take the watermarking system out of the equation. There was AI detection happening even before this. And Google itself has a hard stance on AI.
Look at what it said in the Google Search Central blog, in its guidance on using generative AI content on your website. It’s flagging scaled content abuse when it perceives:
❌ What Google Flags
- Little to no effort
- Little to no originality
- Little to no added value
So watermarking or not, if you’re using an AI system or tool to generate your content, that’s little to no effort. And since AI writes at the average, we’re hitting that little to no originality flag. If both of those things are true, you’re adding little to no value.
It’s why at Sixth City we have taken such a hard stance on this, and we went human, even in the age of AI. We do use it for improvements and efficiency. We use it to enhance human content, and I’m going to get into a little bit more about that in a second.
“Mount AI” and What Happens After the Spike
Let’s look at the problems that AI content has had – once again, even before this watermarking thing came into play. It’s very well known on LinkedIn: these traffic charts that people call Mount AI. What websites were doing was going crazy with these new AI tools and scaling up content massively.
And initially you would see a benefit. You’d see some traffic gains from that. But after not too long, you’d hit a wall, and all of that would fall off a cliff.
A Real Example: Momcozy
There’s a real example of this. This is Momcozy, and they’re pretty authoritative – clearly they’ve done a lot of work to build that authority over time. And it’s just kind of sad to see. On the left, you see sort of natural and really strong growth, and then right about this point here, March 2026, things look a little more aggressive.
We know that they did use AI to scale up their content, and not long after that, things are falling off a cliff. So you have to be really careful with this sort of thing.
The Number I’d Give You: 75% Human, 25% AI
I did an entire other video and wrote a blog on this. It’s called “How Much of Your Content Can AI Write?” and I think it’s really pertinent to this new announcement from Anthropic and their watermarking system. I took a hard stance on this, and I gave you an exact number.
✅ The 75/25 Rule
Aim for 75% human, 25% AI. Human words never enter the candidate set, so they can never be watermarked.
If you’re writing human, then your words are not going to be part of that candidate set. So you’re not going to meet those probabilities for watermarking, and your content is going to be totally clear. Not only that, but it’s going to be durable no matter what system pops up in the future.
What Should That 25% AI Look Like?
Because I’m really specific there as well. I think it’s totally fine to write things like metadata – your title tags and meta descriptions – with AI.
✅ Fair Game for AI
- Metadata: title tags and meta descriptions
- Headings
- Bullet points
- Tables
- Key takeaways
- Notable quotes
That’s where the AI system can kind of search through your content and look for opportunities to enhance it and make it a better user experience. But it’s not acting like a writer.
❌ Keep This Human
Everything else – the entire body of the content – should be human-written. And you should be letting your expertise, your stories, your opinions, and your recommendations come through.
That’s what’s going to make your content hold up over watermarking systems, algorithm updates, whatever. It’s going to help differentiate it. It’s going to be super original, it’s going to rank better, and it’s going to be a much more likely candidate for selection in an AI search result as well.
The Rule of Thumb
💡 Key Rule
Once it’s more than a sentence or two, a human should be writing it.
When I did that video, I shared that rule of thumb. And this whole topic of watermarking – and how that system works – actually makes the advice pretty sound, because we know that longer passages are better for the watermarking system to analyze and detect.
That’s why our writers and our SEOs, when they’re working with clients, don’t let AI write more than a couple of sentences. We really try to stick with these rules; we let AI do its thing in those areas, and then we go human for everything else.
Worried About the Content You’ve Already Published?
So, a really interesting update from Anthropic. Now, if you’ve been using AI to write or working with an agency that has been doing the same, reach out to us.
Our AI SEO agency is happy to give you a free website assessment and discuss how you can align your strategy for future growth and avoid trouble, no matter what updates are coming.
