I've been sitting on Anthropic's watermarking announcement for a little while.
Not because it isn't interesting. It is. Anthropic has developed a clever way to watermark text generated by Claude in a way that can survive copying, pasting, and some amount of editing. They're doing it largely because of transparency requirements coming out of the EU, and to Anthropic's credit, they've also been fairly explicit about what the watermark does and does not prove.
What it proves, broadly speaking, is that Claude was probably involved in producing the text.
And I just could not bring myself to care very much.
I try not to write simply because something happened. Nobody is reading this site as their primary source of breaking AI news, and if you are, then I apologize for being late. I usually wait until I feel like I have something to add beyond pointing at a thing and saying, "Hey, this exists."
What finally made this one interesting to me is that the watermark itself is not really the problem. Anthropic isn't the problem either. They're building something clever to satisfy a requirement that, on its face, makes sense.
The problem is that we're measuring the wrong thing.
Congratulations. AI Touched It.
What I actually care about is not who arranged the words. I care about who did the thinking.
Those two things used to be correlated closely enough that we could mostly treat them as the same question. If you handed me an essay twenty years ago, it was generally reasonable to assume that the person whose name was on it also did most of the thinking required to produce it. Generative AI has broken that assumption, but I think we're making a mistake by assuming it broke it in only one direction.
You see, someone can absolutely ask an AI to generate an argument about a subject they know nothing about, copy the response, put their name on it, and manufacture the appearance of authority. That happens constantly. It sucks. And it makes people suspicious of generative AI.
But someone can also spend hours investigating a subject, develop an original thesis, argue through it with an AI, challenge their own assumptions, refine the idea, change their mind, and then use the model to help turn all of that into coherent prose.
The watermark may look exactly the same in both cases. The intellectual process could not be more different.
The Boat Is On Fire. Apparently That's Good.
There is an old AI training example that I keep coming back to. An agent is trained to play a boat racing game. From our perspective, the goal is obvious: race the boat well and win.
But that isn't actually the goal we gave the machine. How do you tell something with math to "win the race?" Possible, sure. But it is easier to just give it a number... a score... and tell it to make that score as big as possible.
And it sets to work to do just that... the agent eventually discovers a bizarre little corner of the game where it can repeatedly crash, catch fire, collect power-ups, and accumulate more points than it could by actually racing. According to what we wanted it has completely failed. According to what we told it, it has done exactly what we asked. And it is living its best life as a result.
"Maximize the score" and "become good at the game" are adjacent goals, but they are not the same goal.
I think we are doing something similar with AI-generated content. What we really want to know is whether there is a person behind the work who understands the ideas being presented, believes them, can defend them, and is willing to take responsibility for them. That is an extremely difficult thing to measure, so we substitute something easier: was AI involved in generating these words?
That one can apparently be watermarked. And as clever as the solution is... it is nothing more than a boat rampaging in circles, catching fire, and maximizing its score. Interesting for all the wrong reasons.
I Expect My Junior Developers to Use AI
This becomes especially obvious to me in software development. If a junior developer submits a pull request today, I assume they used AI. In fact, I hope they used quite a lot of it. If they didn't, then they won't last long in software dev the way things are going today. Indeed, if I discovered that one of my developers had spent four hours manually doing something Claude or Codex could have helped them accomplish in forty minutes because they thought using AI was somehow cheating, I would not admire their work ethic. I would question their judgment... and perhaps their sanity.
The meaningful review happens after the code exists. Does the developer understand what the implementation is doing? Can they explain why this approach makes sense, where it can fail, and what tradeoffs were made? If I introduce a new constraint, can they reason through how the solution changes?
Most importantly, do they own the pull request?
AI can generate code. It cannot transfer accountability.
If you are asking me to merge something into a production system, you are telling me that you understand it well enough to put your name behind it. Whether your fingers personally typed every character is almost completely irrelevant to me.
The same principle applies to writing, although there is another dimension there that matters too: taste.
AI Has Taste. That's Part of the Problem.
I was giving an AI workshop recently and had a group of teachers use AI to write an email addressing a behavioral problem with a student. The first teacher generated hers and was immediately impressed. It was polished, professional, empathetic, and clear. She said she could probably send it as-is.
I told her to hold off for a minute.
Another teacher generated one. Then another. Then another. Eventually we started reading them out loud.
The first teacher, the one who had been most excited to immediately send hers, noticed the problem first.
"Why do all of these sound exactly alike?"
And there it is... That is the taste problem.
AI clearly has taste. If it didn't, its writing would not be so recognizable. Its default taste is also pretty good. Hell, it is objectively, mathematically good. That is part of the reason people are impressed when they first use it... and the main problem. Math ruins everything.
Read one example in isolation and it can seem excellent. Read twenty examples in a row and the sameness becomes impossible to ignore. The transitions feel familiar. The cadence feels familiar. The rhetorical structure feels... you guessed it: familiar. The same little writing habits show up again and again until the work starts feeling less like an individual voice and more like one person has been hired to ghostwrite for the entire internet.
But... and here's the rub... that does not mean the underlying thought is bad.
Taste and Ownership Are Not the Same Thing
Imagine someone who has spent their entire career becoming deeply knowledgeable about a subject but does not speak English particularly well. They have an original idea. They understand it completely. They can defend it. They developed it themselves. Then they use AI to express that idea fluently in English.
Did AI help write it? Absolutely.
Did AI think it? No.
The same is true of someone who is simply not a very good writer. There are brilliant people who cannot produce an engaging paragraph to save their lives. Writing ability and thinking ability overlap, but they are obviously not identical.
Generative AI can give those people a voice they previously did not have with certain audiences. I do not want to throw that away simply because the resulting prose occasionally sounds a little too much like Claude.
If the idea underneath it is genuinely interesting, I can forgive quite a lot in the presentation.
You Still Have to Bring Something
The opposite is also true. We have all encountered the YouTube video, LinkedIn article, blog post, or presentation where it becomes painfully obvious that someone prompted an AI and then simply read whatever came back.
At some point, the question becomes: why am I listening to you?
Not because you used AI. But because you did not add anything.
If I could have typed the same prompt myself and gotten essentially the same answer, then what exactly is the value being contributed by the person presenting it? The scarce thing was never the words. The scarce thing was the judgment, experience, context, and perspective that made those words worth hearing from that person in the first place.
That is actually why it took me a while to write about Anthropic's watermark.
I knew about it. I understood the basic idea. I could have asked an AI the day it was announced to write me a thoughtful article about the implications of AI watermarking, and I am sure it would have produced something perfectly respectable.
That is not why I write.
I write when I think I have something to add.
Anthropic Isn't the Villain Here
In this case, I did not have much interest in manufacturing outrage at Anthropic. They're complying with regulation, and they're being fairly transparent about both why they're doing it and the limitations of what they've built.
The EU is not really the villain here either. The underlying concern is legitimate. Deception, impersonation, misinformation, fraud, and synthetic content at massive scale are all real problems, and governments have every reason to care about them.
I respect that the EU is at least trying. The difficulty is that governing bodies are built to create stable rules, while AI is changing at a pace that makes almost every durable assumption feel temporary.
We are trying to establish social and regulatory norms around a technology whose capabilities and use cases are changing much faster than institutions are designed to move. So we reach for things we can measure: labels, watermarks, disclosure requirements, technical fingerprints.
Those things may be useful in some contexts, especially when deception itself is the goal. If someone is impersonating a public figure, fabricating evidence, or running automated influence campaigns, provenance matters quite a lot.
But we should be careful not to confuse the measurable thing with the important thing.
The Ground Is Moving Too Fast
The way ideas are created and communicated has already changed permanently, and writing is only the most obvious example.
The same shift is happening in software, design, images, research, science, analysis, and almost every other field where language can be used to describe an outcome and tools can be attached to a system capable of acting on that description.
That is the thing about language. It is meant to describe almost anything.
Build a system that becomes extraordinarily capable with language, give it tools that allow language to affect the world, and the distinction between "language model" and "general-purpose tool for doing things" starts getting very fuzzy.
Pandora's box is open. We are not going back.
The problem is that our rules are being written while the meaning of the categories themselves is changing underneath us. "AI-generated" already means something radically different than it did a couple of years ago, and it will probably mean something different again a couple of years from now.
That makes regulating the mechanism incredibly difficult.
If It Moves You, Does It Matter if AI Was Involved?
I do not think the useful question going forward is whether AI participated. Increasingly, that will be like asking whether someone used a computer. The harder question is whether there is a human being who actually owns the result. Do they understand what is being said, can they defend it, can they recognize when it is wrong, and are they willing to take responsibility for it?
Those are frustrating questions because they do not fit neatly into metadata. You cannot cryptographically fingerprint them. You actually have to engage with the person and the work.
Which is inconvenient.
But inconvenience does not make the easier question meaningful.
Because if you loved a thing and then dismissed it when you found out AI had a part in it... you're not taking a moral stand, or at least not the one you think you are. You're just being closed minded.
But yeah... Anthropic built a genuinely clever watermark. It can tell us that Claude probably touched something.
Cool.
Now we still have to figure out the part that actually matters.