I write for a living. Watermarking AI text is missing the entire point of the AI slop debate

I am not easily annoyed by people devaluing the written word, even though it’s fundamental to my career. Language is fluid. It evolves. I have no more right to choose how it does so than anyone else. 

I was surprised then to find myself as disappointed as I was reading Anthropic’s recent response to the EU’s new AI Act. It involves watermarking the text its AI model, Claude, produces, anywhere in the world and in perpetuity. Reaction has been mixed: people appreciate transparency, but not everyone wants it to come at the cost of quality and intent. Veteran tech blogger John Gruber went so far as to call it a “perversion of writing.” 

He’s not not right.

Why did it upset me, when the intent is clearly to preserve the integrity of the human written word? Ultimately, this is what we offer as a service to our clients.

Then I realised: it’s because it is an affront to the reader as well as the writer. AI watermarking in this way asks you to give up on your judgment, in the same way relying on generative AI does for the author.

The internet is now a mess: somewhere between 10% and 75% of new content is AI generated. Yes. It’s trivial to one-click blogspam.

But when anyone can churn out words for nothing, words stop being worth much on their own. What’s scarce now is credibility: someone worth listening to. So instead of arguing about how much AI was involved in making something, we need to get back to asking whether we trust the source, and whether what we read was useful or interesting. Provenance worked well enough before. It matters even more now.

Let me explain.

Steganography: big word for a blunt tool

In short, the EU AI Act requires output from AI models within the EU to be watermarked. I’m not concerned about this for images and videos, since these can be embedded in the metadata. Transparency is good. You can already check some files here. Carry on. 

But Anthropic’s solution to do the same for text Claude creates, that you can copy, and paste, and edit, involves it adulterating the actual word choices it makes (Read about the concept of steganography here; Google has also implemented this approach in Gemini, while OpenAI has just announced similar plans, but is quite clearly far less enthusiastic about it and is restricting the change to users inside the EU). With the right algorithm or key (which Anthropic is withholding for now), and a long enough body of text, you can then assess how likely it was that Claude created the copy. In theory.

Conceptually, this is problematic. Instead of you paying Anthropic for its best attempt at conveying meaning for you at speed, it is now, apparently, not going to. And not just in the EU: this is being applied globally. I asked Claude to stitch together several piecemeal transcripts of the same long meeting for me last month, which was helpful, and, I think, benign. Is it actually going to change the words that people said in the output from now on to meet these requirements? 

It’s also trivially easy to work around, which bad actors will (Spoiler: use a different AI). It’s the sheer number of false positives and the pointless drama that will ensue that concerns me most though.

Get ready for the AI Gotcha Era

Here at Archetype, we have a transparent AI policy in place we share with clients. We’re not in the business of fobbing people off with AI content they do not want. But with text watermarking in play, I can still foresee plenty of either acceptable or unavoidable ways that any business can end up putting words out into the world which could be deemed as touched by Claude at some point.

Say, if you were to use an AI tool to put a lovingly crafted article into a specific template, because AI is perfectly capable of doing so now and, really, life is too short to be doing the same task over and over if you don’t need to. That may now be branded as Claude-generated, but the creative labour was still all human.

Or, suppose some draft copy supplied by a third party (like a quote from a client’s customer) was unknowingly generated by Claude. Steganography doesn’t provide a way to assess if Claude created watermarked text from scratch, or just proofed it in a conversation.

I could go on. This strategy is as helpful as the lie detector tests on the Jeremy Kyle Show, which were only slightly more accurate than flipping a coin. 

The provenance problem

I am the first person to defend the importance of expert craft in editorial, but this approach to watermarking is a distraction. It’s a technical attempt to solve a sociological problem, and like most technical solutions to social problems, it is a bungled one.

And not just because Anthropic hasn’t actually released the keys to decrypt the watermark it is using, or said when it will or who to (OpenAI is sharing its own with specific researchers to start with). By focusing on the “how” of the text – the invisible, algorithmic fingerprints left by a machine – we are missing the “who” and the “why”. We are spending our collective cognitive energy trying to build a communal polygraph when we should be reinvesting in the art of discernment, something we’ve spent centuries refining but now seem happy to set aside.

We have been dealing with this problem since long before the printing press, after all. Before we had ink, we had reputation. Before we had algorithms, we had provenance. Deciding to trust the source, not the format. You apply that to the New York Times, any YouTuber or some bloke on an orange box in Hyde Park, subconsciously, every day.

You can’t automate away trust. But you should be allowed to run an article past a house style guide skill in an AI tool to see if you have missed anything, without damaging that.

And you should definitely be able to do that without having to deal with the irony of others leaning on an AI solution to spot AI copy from human copy, and just binning off their own human intuition.

The human variable

When you read something, your brain performs a sophisticated, subconscious audit. You aren’t looking for steganographic patterns; you are looking for a signal in the noise. 

You are asking “Do I rate this person? Do they have a track record of being right? Are they worth listening to?” 

If the answer to any of those is “no”, it doesn’t matter if the text was written by a Nobel laureate or a loosely-prompted LLM. You’re going to lose interest and do something else.

This is where AI-shortcuts stop leading to great content marketing. AI can tell you what a CEO might say about market volatility based on a thousand existing whitepapers. But an AI will never go out, find the most interesting, slightly eccentric engineer in your R&D labs, listen to their actual lived struggle, and translate that truth into a narrative that makes a reader’s skin prickle – and maybe move them down the marketing funnel. 

AI can’t find the amazing people with stories worth telling. It can’t get their buy-in. It can’t build the relationship required to extract the “why” behind the “what”.

The future of communication isn’t about detecting the machine, or at least it shouldn’t be. Once the inevitable, silly game of media Gotcha this will cause has died down, we’ll come to realise it’s about doubling down on the human touch. It’s about finding the voices that are too interesting, too reliable, and too insightful to be mistaken for a statistical probability.

Funnily enough, that’s how AI engines work too. When ChatGPT or Claude decide who to quote, they aren’t hunting for watermarks. They’re weighing the source: who said it, what they’ve got to back it up, and who else rates them. In other words, GEO and AEO reward much the same things human readers do: credible sources, identifiable expertise and a story that holds together across different places. It’s why Trust is the T in VISTA, the framework we use to help brands get seen and cited by people and by AI.

If you need help with that, contact Archetype.

PS This article isn’t watermarked xx