This week, Anthropic announced that, in line with the EU AI Act’s Transparency Code, it will soon begin to watermark content generated by its products so readers globally will know what came from a machine and what was actually a DIY job. 

A few months ago, I was sitting on the bench at our farm, looking out over the hills, huddled by the heater on the veranda to keep warm in single-digit temperatures. In that moment, with no reception, no texts coming in, no social media to scroll, my mind had a chance to disconnect… and I got inspired. 

That day, using my phone as a keypad only, I wrote four blogs in quick succession, all by myself. And I loved every minute of it.

I guess that’s probably why I’m ever so slightly optimistic about Anthropic’s latest announcement, that soon, its AI-generated content will be watermarked. 

My wish, probably naively, is that the mark will correct a path we’ve been stumbling down too quickly, and help us pivot back to thinking more for ourselves. I hope it sees us writing more for ourselves, and of course, reading more that is coming from the minds, mouths and fingers of other human beings. 

AI just makes too many things easy. 

It takes the 60+ minutes you might have spent writing a blog, and condenses that effort into 90 seconds. It takes the hours you would have lost researching, and delivers you a smorgasbord of information for your choosing. It generates a month’s worth of post content (albeit pretty average) in a minute.

And not all of that is a bad thing. 

But when it produces a song that hits the charts, challenging the work of artists who have spilled their stories onto a smudged page during a coffee-fuelled creative session, it isn’t great. 

When it takes a three-sentence prompt, and turns it into a ‘personal’ story that people invest in, relate to, take cues from, rather than drawing on their own experience and that of other real people, that can be concerning. 

When it creates content for blogs, or posts, or schoolwork, or conversations that starts with ‘I think’, and someone will own and claim that opinion without having fully considered their ‘new position’ on the topic at hand, it becomes potentially dangerous. 

I love stories, songs, and hearing the diverse, interesting, controversial, genius, silly, hilarious opinions of other people on all sorts of issues. Because whether I tap a foot or not, whether I agree or disagree with that opinion, it has been shaped by someone’s human experience.

So, what does the watermark mean?

This new code is a regulatory response to the rapid pace at which AI is evolving, but I think (there I go again!) it’s also a reaction to the exhaustion so many of us are feeling at continuously swallowing AI slop.

We are media beings: we love our internet, our socials, our stories, and so many people are just tired of wondering ‘is that real?’ or knowing immediately that it isn’t. 

Right now, what will happen when this watermark rolls out is speculation. 

If you hang out on Reddit, you’ll see predictions of mass exodus from Anthropic platforms, but with the widespread reach of the AI Act, that seems futile. Given others have already sourced or built watermarking tech, next week it could be any other platform making a similar announcement.

Others are predicting that it will make no difference. This is one of the most interesting opinions, because it indicates how much we have psychologically adapted or accepted, since AI took off.

The concept assumes eventually all of us average Joes will be able to see marked content, and though we are sick of AI, we’ll keep consuming it… Maybe slightly frustrated and we might complain about it a bit, but nevertheless we’ll still stick our fork in and take another big bite. 

There’s a whole truckload of scientific opinion (and early research) about how AI has already affected our brains and psychology, and a lot more planned for how it will continue to do so going forward. 

What I want to see, is that this code creates a line in the sand about what people will happily consume from AI and what they won’t, and in doing so, changes the behaviour of all of us who do make any sort of content, whether it’s blogs and posts, or novels and songs. 

I hope that we start to more closely evaluate what we feel we need to experience genuinely from other people, and what can be machine-generated because it won’t greatly affect our humanity and how that evolves.

Perhaps we’ll be comfortable with it being used fairly profusely in most workplaces, but vote with our feet (or clicks or eyes) when it comes to poetry, songs, opinions, stories, shared experiences, art. 

By the end of the year, maybe we’ll know the answer.

How will it work?

This is a really good question, and at the moment, the details are scarce. 

What is known is that this watermark is not going to be like the shiny one we have on our licences to tell a real from a fake, nor will it be that big ‘draft’ stamp so common in early versions of Microsoft Word documents. 

This watermark, by all reports, is embedded into the pattern of words generated by the AI platform. 

For most of us to begin with, though we can see signs of AI in text, we’ll be largely blind to it. The text generated will seem normal, the LLM will still do what you asked it to do, it will select words that feel ‘natural’ and produce the report or post or blog or answer you need. But hidden in there, in that word selection, is a sign that this compelling content was indeed passed through Claude. 

So, what if you copy the text into notes then into your own Word doc? Can you remove the watermark? According to answers so far, it will likely have no effect. The watermark will be carried when copied and pasted.

What if you read it on screen then type it up yourself? Again, apparently, a no-go, because it’s the very act of using those exact same words that will get you. 

What if you tweak the text a little and remove the pattern? Good luck… by all accounts this isn’t a ‘trick word’, it’s much more than that. And though it was acknowledged that changing the copy ‘enough’, or producing content in very small word sets may remove the watermark, the boundaries aren’t yet known. 

Beyond that, for those who do write from scratch, if AI is then used to proof-read, the writer runs the risk of also having their document, which was written by human hands, watermarked ‘AI’ if changes are applied within the tool. Anthropic has reportedly been open about the fact that the tool therefore doesn’t prove AI authored a particular piece of writing, just that the piece passed through it in some form, at some point.

According to Anthropic, all models released after 2 August 2026 will automatically have the watermark tech, and since most current models were pre that date, a lot of users won’t be immediately affected by it. But… and there is a but, they will also be retrofitting it to all existing models, so that exemption may last only for a very limited time.

It will apply to AI-generated text, and to files like SVG, PNG, and JPG (for the latter, C2PA open standard is reportedly being used – interesting discussions already happening there about workarounds). 

Notable is that Anthropic hasn’t announced when any kind of detection tool will be available, or what its margin of error will be. Looking at current AI detectors on the market, false positives can be quite common. Will this have the same issue?

Perhaps most intriguing, is the arms race this creates among those working to build tools that can remove the watermark, a concept that those in-the-know seemingly say, will be possible. 

So, while people from all walks of life (including those in professional spaces contemplating the impact on areas like SEO and AEO) are on edge about the Anthropic watermark, after all of that, there’s a chance the fuss will be about nothing.