This Article Used AI. Does That Make It Less Mine?

I’m broadly in favour of knowing where things come from. As AI-generated content becomes increasingly difficult to distinguish from human-generated content, provenance matters. If an image has been generated or altered by AI, there are circumstances where I absolutely want to know that. The same applies to text. So I understand the thinking behind Anthropic introducing watermarking into Claude, and at a basic level I think greater transparency about how something was produced is a good thing.

What interests me more is what happens next. Because I can easily imagine us moving from “AI was involved in creating this” to “this was written by AI”, and from there to “therefore this isn’t really the person’s work”. Those three statements sound superficially similar, but they mean very different things. As AI becomes embedded in everyday knowledge work, I think that distinction is going to matter a lot.

Consider something as ordinary as writing a job application. I could write the whole thing myself and ask Claude to check the spelling. I could ask it to make my writing clearer. I could give it everything I know about my career and the role, spend an hour discussing which experiences are relevant, challenge some of its suggestions, work out the argument I want to make and then ask it to help draft the final version. Or I could simply paste in the job description and say, “Write me a brilliant application for this.” Those are clearly not equivalent activities, even though AI was involved in all of them.

That is where watermarking becomes interesting. Anthropic itself is quite careful about this distinction: its watermark can indicate that Claude was involved with a piece of text, but it cannot tell you exactly what that involvement represented. Claude might have originated the text, substantially edited it, translated it or otherwise transformed something whose underlying ideas came from a human. The watermark gives us useful information about provenance, but it doesn’t give us a verdict on authorship.

I worry that we might nevertheless start treating it as one.

We need something more sophisticated than human or AI

Very little professional work is created entirely without tools or other people influencing it. If I use spellcheck on this article, nobody concludes that I didn’t write it. If somebody reads a draft and tells me the argument doesn’t work, and I rewrite it as a result, we don’t normally transfer authorship to them. Researchers use search engines, analysts use software, writers have editors and executives have teams who help turn ideas into documents.

AI is obviously different. It would be disingenuous to pretend that Claude generating three paragraphs is equivalent to Word correcting a spelling mistake. AI can contribute language, structure, analysis and even ideas to a degree that previous tools generally could not. But that is precisely why I think a simple distinction between “human” and “AI” is becoming less useful. There is now an enormous spectrum between writing every word yourself and contributing almost nothing to the finished product.

For some activities, that spectrum matters enormously. If I am sitting an examination designed to establish what I personally know, using AI to produce my answer clearly undermines what is being assessed. There will be professional settings where disclosure is essential, and others where AI use may simply be inappropriate. Provenance is useful partly because context matters.

But take recruitment. If I am recruiting somebody into a senior role, I’m not sure I care very much whether Claude helped them improve the wording of their CV. I care whether the experience described is actually theirs, whether they understand what they claim to have done, whether they can demonstrate the judgement the role requires and whether they can have a credible conversation about it. In fact, if using AI effectively is going to be part of how they perform the job, insisting that they demonstrate their suitability for it without using AI starts to become a slightly odd test.

The same applies to a strategy, a report or an article like this. If somebody uses AI to challenge an argument, identify a weakness, research an issue, improve the structure or help express something more clearly, I don’t automatically think the resulting work becomes less authentic. If they type one sentence into a chatbot and publish whatever comes back without reading it, I probably do. The interesting question isn’t simply whether a tool touched the work. It is what contribution the person actually made, what they understand and what they are prepared to stand behind.

The bigger danger is what we do with the watermark

This is where I think the unintended consequences could become more important than the technology itself. Once something is machine-readable, organisations naturally start building processes around it. It isn’t difficult to imagine an applicant tracking system flagging or downgrading an AI-watermarked application, an educational institution treating every marked passage as suspicious, or an organisation applying additional scrutiny to any document in which AI involvement can be detected.

Those rules would be wonderfully easy to implement. They would also collapse a complicated question about provenance, judgement and human contribution into a binary one: AI involved, yes or no.

Worse, they could create exactly the wrong incentive. Anthropic acknowledges that sufficiently extensive rewriting can remove the detectable text watermark, and the absence of a watermark cannot prove that AI wasn’t involved in producing something. If transparent AI use attracts a penalty, people won’t necessarily stop using AI. Some will simply become better at hiding the fact that they used it. We could end up penalising the person who openly uses AI as a thinking and editing partner while rewarding the person who uses it extensively and then successfully removes the evidence.

That seems a particularly poor outcome at a time when organisations are simultaneously encouraging people to become more capable users of AI.

There is a wider question here about what we actually mean by authorship. Historically, authorship and execution were closely connected. If my name appeared on an article, there was a reasonable assumption that I had personally written most of its words. If I produced a piece of analysis, doing the analysis was part of the evidence that I understood it. AI is beginning to separate those things. Increasingly, someone can be responsible for the direction, reasoning and decisions behind a piece of work without personally executing every step involved in producing it.

I don’t think that makes authorship meaningless. But it may mean we need a more useful definition of it. When I say “this is my work”, perhaps what increasingly matters is that these are ideas I believe, that I understand the argument, that I have challenged and checked what has been produced, that I made the important decisions and that I am prepared to stand behind the result. AI may have played a significant part in getting me there.

That feels much closer to how I already think about using AI in other areas. I can build software now that is beyond what I could independently code from scratch. That doesn’t make me a better software engineer than someone who could. But it does mean I can turn more of my ideas into working things. The interesting question is not whether AI was involved; it is what I contributed, what I understand, where my limitations are and whether the resulting thing is fit for purpose.

The same principle increasingly applies to knowledge work.

Provenance is information, not a judgement

This is why I think watermarking could be valuable. Knowing that AI was involved gives us another piece of information about how something came into existence. In some contexts that information will be extremely important. In others it may barely matter.

What I don’t think the watermark can tell us is whether something is good, whether its ideas are original, whether the person whose name is attached to it understands it, or whether they should receive credit for it. Those require judgement, context and, increasingly, a better understanding of the relationship between the human and the tools they used.

Perhaps that is the adjustment we need to make. Instead of asking simply, “Was this made with AI?”, we need to become much better at asking “How was AI used, what did the person contribute, and what are we actually trying to assess?”

That is harder. It is certainly harder to automate. But it also feels much closer to the world we are heading into.

Because if AI really does become embedded in how most of us work, “AI was involved” will eventually tell us remarkably little.

What will matter is what the human brought to the process, what they understand, and whether they are prepared to stand behind the result.