Anthropic recently announced that future Claude models will watermark generated text so it is possible to estimate whether Claude was involved in creating it.

That caught my attention for a few different reasons.

I spent years working in print design, where watermarks are associated with provenance and authenticity. It also immediately took me back to my university dissertation, What’s My Name?, an exploration of the origins of graffiti and the human instinct to make a mark.

I was here.

There is something strangely familiar about an AI system leaving its own invisible signature through a piece of writing. But the more I read about Claude’s watermark, the less interested I became in the watermark itself.

The interesting question is what it cannot tell us.

Claude isn’t hiding secret characters in your copy

My first instinct was probably the obvious one: perhaps the model simply inserts hidden characters or a detectable pattern into its output. That isn’t how Anthropic says it works.

Claude’s watermark uses subtle patterns in the word choices the model makes while generating text. When there are several equally reasonable ways to continue a sentence, the watermark influences those otherwise low-stakes choices in a way that can later be statistically detected by someone with the appropriate key. Nothing visible is added to the text. There are no secret characters, and a reader should not be able to tell whether the watermark is present simply by reading the copy.

There are limitations too. Short passages provide less evidence, factual writing offers fewer opportunities for the system to vary its word choices, and sufficiently heavy rewriting can weaken or remove the watermark. Most importantly, Anthropic makes a distinction that sits at the centre of this whole subject:

A watermark might indicate that Claude was involved.

It cannot tell whether Claude wrote the piece or heavily edited something a person had already written.

That changes the question considerably.

AI involvement isn’t the same as AI authorship

There is a huge difference between opening an LLM and asking:

“Write me a thought-leadership article about marketing.

…and

arriving with an observation from your own work, evidence, experience, something you disagree with and a conclusion you are prepared to defend, then using AI to research, challenge, structure and improve how you communicate it.

Both involve AI.

But are they really the same kind of authorship?

I don’t think they are.

This matters to me because I use AI extensively in my own writing.

I rarely arrive with a neat 1,000-word article ready to polish. More often I arrive with too much: observations, half-formed arguments, examples, questions, things I have noticed at work and sometimes things I am not yet sure I agree with myself.

AI helps me interrogate that material.

• What is the strongest argument?
• Where am I making an assumption?
• What evidence supports it?
• What contradicts it?
• Am I trying to squeeze three subjects into one article?
• What should I remove?

Sometimes the model suggests something better. Sometimes it produces something that sounds entirely plausible and I disagree with it immediately. That disagreement is often useful too. Eventually I end up with something that represents what I actually think considerably better than the pile of notes I started with.

So does that make the finished article less mine?

At what point would it?

50%? 80%? The first draft?

The more I think about it, the less useful percentage-based definitions of authorship seem. Imagine I supply the entire argument, all of the examples and the conclusion, but AI rewrites 80% of my sentences.

Now imagine somebody else types a ten-word prompt and accepts 95% of what the model generates.

The first article may contain far more AI-produced language. The second may contain far less human thought.

Counting words tells us very little about the intellectual contribution.

A watermark might therefore help establish the provenance of the words.

It still cannot establish the originality of the thinking behind them.

Those are different questions.

And I think we are going to need to become much better at separating them.

So what did the human actually contribute?

The test I am increasingly drawn towards is this:

What did the author contribute that the model could not have produced without them?

The customer conversation.

The proprietary data.

The mistake.

The experiment that failed.

The disagreement.

A first-hand observation.

Twenty years of pattern recognition.

The knowledge that something technically correct is commercially irrelevant.

The judgement to reject a plausible answer because it is wrong in this particular context.

None of this means AI is incapable of generating an interesting idea. It can.

An LLM can generate hypotheses, challenge conventional wisdom, connect concepts and propose things I had not considered.

But there is a difference between generating a plausible hypothesis and bringing new information from the world into the conversation.

If a customer told me something unusual yesterday, the model does not know that unless I tell it.

Once I do, however, AI can become exceptionally useful again.

It can test the observation against published evidence, look for counterarguments, identify weaknesses and help me work out whether I have discovered something interesting or merely an anecdote.

That is the relationship with AI I find much more valuable.

Not replacing the thinking.

Making the thinking work harder.

A five-question test for AI-assisted authorship

Rather than trying to calculate what percentage of an article came from a human, these are the questions I am starting to find more useful:

1. Where did the original observation come from?

Was there something the author genuinely noticed, experienced, measured or questioned?

2. What information did the human add?

Experience, proprietary data, customer conversations and professional judgement all change the information available to the model.

3. Was the model allowed to manufacture the argument as well as the words?

There is nothing inherently wrong with using AI to generate ideas, but there is a difference between exploring possibilities and attaching your name to an opinion you never really formed.

4. Did the person challenge the output?

AI can produce exceptionally convincing wrong answers. Knowing when something is incomplete, badly prioritised or simply wrong is part of the intellectual work.

5. Will the author take responsibility for the conclusion?

If somebody challenges the argument, can the named author explain why they believe it?

For me, that final question matters enormously.

Authorship isn’t simply about who moved the keyboard.

It is also about who owns the judgement.

“AI slop” is really an information problem

The internet is already filling with material that looks professionally written but says very little.

What frustrates me isn’t AI use.

It is outsourcing both the thinking and the writing, then presenting the result as expertise.

LinkedIn has started talking openly about low-effort AI content: material that may appear polished but lacks a clear perspective, unique experience or substance.

Interestingly, its guidance isn’t simply telling professionals to avoid AI.

AI can be useful for articulating ideas, refining language and making writing more concise.

The distinction that matters is value.

AI-assisted content can still reflect a real person’s perspective, experience or expertise.

Generic, repetitive content without a meaningful contribution is something very different.

That feels much closer to the distinction I care about.

The problem isn’t necessarily that AI touched the content.

The problem is when the human added nothing worth preserving.

What about Google and search?

There is a similar misconception around search.

It is worth clarifying first that browsers such as Chrome or Safari do not decide which pages rank. Search engines do.

And Google’s position isn’t simply:

“AI content bad. Human content good.”

Generative AI can be useful for researching a subject and helping structure original content.

The problem comes when businesses generate large volumes of material without adding meaningful value for the user.

That distinction is important for marketers.

If every business has access to broadly the same models, and everyone asks those models to research the same publicly available information and manufacture an article from it, producing more content does not necessarily create more authority.

It may simply create more versions of the same answer.

The interesting distinction isn’t human versus AI.

It’s original value versus commodity output.

Can you identify AI writing by its style?

Probably not reliably.

People increasingly point to particular sentence structures, excessively tidy lists, certain phrases or even things like long dashes as evidence that something was written by ChatGPT or another LLM. They can certainly be clues but they are not proof.

Humans use those conventions too. Models can be prompted not to use them. Editing can remove them in seconds.

Ironically, trying to identify AI authorship purely from writing style is exactly the problem watermarking is attempting to address.

But even a technically reliable watermark brings us back to the same limitation.

It tells us something about production.

Not necessarily about contribution.

This becomes a question of professional integrity

I am not trying to answer the legal question of who owns copyright in AI-assisted work here. I am much more interested in the professional one. If I put my name against an article, what am I claiming?

For me, I am claiming:

• That the argument represents what I believe.
• That I have questioned it.
• That I have checked it.
• That the experiences attributed to me are mine.
• That I have not allowed an LLM to manufacture expertise I do not possess.
• That if somebody challenges the conclusion, I am prepared to discuss and defend it.

AI can help me express that argument.

Sometimes it can help me improve it considerably.

That doesn’t trouble me.

Presenting generated expertise as personal expertise does.

Where did the thinking come from?

As AI involvement becomes easier to identify, I suspect we are going to spend a lot of time debating labels, watermarks and whether something was “AI-generated”. Those conversations are useful. But I think there is a more important question underneath them.

Not:

“Did AI touch this?”

But:

“Where did the thinking come from?”

A watermark might eventually help us answer the first question with increasing confidence. I am not convinced technology alone can answer the second. And perhaps that is exactly where human authorship still matters.