Claude Is Watermarking AI Content. I Think We’re Asking the Wrong Question.

I’ve seen several posts this week about Claude introducing invisible watermarks into AI-generated text, and I’ll admit that when I first read them, I had one of those moments where I thought, bloody hell, what does this mean now?

The conversation seemed to move very quickly from Claude introducing watermarking, to Google being able to identify AI content, and then to the suggestion that websites using AI-generated content could eventually be penalised in search.

I use AI every day in my business, so naturally I wanted to understand whether this was something I actually needed to be concerned about, rather than getting caught up in another round of AI panic.

I went back and read what Anthropic is actually saying, and then looked at Google's own guidance around AI-generated content.

What I found was considerably more nuanced than some of the posts I’d been reading.

Anthropic says that its newer Claude models will include machine-readable marking in generated text. It also makes an important distinction that seems to be getting lost in some of the conversation. A detected mark can indicate that content has been processed by Claude, which does not necessarily mean Claude originated the content.

Anthropic specifically gives examples of people using Claude to proofread, translate or summarise something they originally created themselves.

That got me thinking, because I’m not sure the term “AI-generated content” is particularly helpful anymore.

If I spend an hour working through an idea based on my own experience, client work and methodology, then use AI to challenge my thinking, help me organise a messy brain dump and turn it into something another person can actually understand, I’ve used AI.

Someone else could type, “Write me 50 SEO articles about AI for small businesses,” publish whatever comes back without adding much of themselves to it, and they’ve used AI too.

Technically, we could both end up being described as producing AI-assisted or AI-generated content, depending on where somebody chooses to draw the line.

Yet what has actually happened in those two situations is completely different.

And this is where I think the watermarking conversation becomes much more interesting than whether Google will eventually be able to detect a particular pattern in some text.

A watermark might tell us that AI has been involved somewhere in the process. It can’t necessarily tell us where the thinking came from, how much expertise went into it, whether the person challenged what the AI produced, or whether the ideas came from twenty years of experience or a thirty-second prompt.

That seems to me to be the much bigger question.

What is Google actually saying?

I went back to Google's own guidance because quite a few of the posts I’ve seen make the jump from AI watermarking to possible SEO penalties.

Google currently says that generative AI can be useful for researching a topic and adding structure to original content. Its concern is the use of generative AI to produce large amounts of content without adding value for users, particularly when that content is being created primarily to manipulate search rankings.

You can read Google’s guidance on using generative AI content here.

That feels like an important distinction, because Google isn't simply saying, “AI touched this, therefore it’s bad.”

Its wider guidance continues to focus on useful, original, people-first content, and Google has also started talking about the importance of creating unique, non-commodity content and bringing first-hand experience and a genuine point of view to what we publish.

You can read Google’s guidance for succeeding in AI search experiences here.

I find that considerably more useful than trying to work out whether every sentence on a website has ever passed through an AI system.

I use AI. A lot.

I don't particularly feel the need to hide that.

I use ChatGPT as a thinking partner. Sometimes I arrive with a fairly developed idea and use it to challenge what I’m thinking. Sometimes I dictate a complete jumble of thoughts and use it to help me work out what I’m actually trying to say. I use it for research, structure, editing, problem-solving and exploring things from different perspectives.

And yes, I use it when I write.

Sometimes it helps enormously. Sometimes it completely misses the point and produces something that sounds nothing like me, and I have to pull it apart and start again.

That process matters.

Because my concern with AI has never really been whether an expert uses it. I’m far more interested in what happens to their expertise when they do.

I’ve seen how easy it is for somebody with years of experience and a very distinctive way of thinking to gradually start producing the same polished, generic language as everybody else. The words might be perfectly competent, but something of the person starts disappearing.

That is one of the reasons I developed EmpowerAi™ as a methodology in the first place. I wanted a structured way of using AI that protects the expertise, judgement, voice, intellectual property and human thinking underneath the technology, rather than simply focusing on how much faster we can produce things.

The watermark conversation actually reinforces that for me.

If AI is going to become increasingly visible within the provenance of our work, then trying to pretend we don’t use it seems like a fairly pointless direction to go in.

I’d rather be able to explain how I use it.

I think we’re going to need a better definition of authorship

This becomes even more complicated as AI gets embedded into the ordinary tools we use every day.

Someone might write something themselves and use AI to proofread it. Another person might use AI to help structure an argument. Someone else might have a long conversation with AI that changes how they think about a subject before they write a single word.

Where exactly does “human-written” stop and “AI-generated” begin?

I’m not sure there is going to be a particularly clean answer to that.

Google has been saying for some time that its systems are focused on the quality and usefulness of content rather than simply how that content was produced. It has also been clear that using automation primarily to manipulate search rankings is where the problem arises.

You can read Google’s position on AI-generated content and Search here.

That seems a much more sensible way of looking at where we're heading.

For me, the interesting question is becoming less about whether AI was involved and more about whether there is still identifiable human expertise underneath what has been produced.

Did the person bring something to the conversation that the AI couldn't have produced without them? Is there experience behind it? Is there judgement? Is there an original methodology, opinion or interpretation? And ultimately, is somebody prepared to put their name to it and take responsibility for it?

Those are the things I care about when I read somebody else's work, and they are the things I want to protect in my own.

So I’ll keep watching what happens with watermarking, particularly if other AI companies start introducing similar systems. But I’m not changing how I work because of it.

I think the more useful conversation is the one underneath all of this.

As AI becomes involved in more of what we create, asking “Was AI used?” is probably going to tell us less and less.

I’m much more interested in asking “Where did the thinking come from?”

Because that tells me something about the work.