Claude wrote this?

Here is how the post you’re reading right now got made. I keep an index of my past writing and a backlog of ideas: things I find myself saying in conversations that never quite made it onto a page. Each week I brainstorm with Claude about which idea is timely/relevant/interesting. Then I have Claude ask me a ton of questions. My answers, typed out in my own words, become the bulk of the post. Claude helps assemble and shape, and then I edit. Heavily. I rewrite by hand, I fact check, I cut, I rephrase, I read it again in the morning and cut some more. Every idea is mine, most sentences were directly pulled from my answers, and every word ships because I chose to keep it.

I’m telling you this because Anthropic just announced that, as far as anyone will ever be able to tell, none of that matters.

What they announced

Last week Anthropic said that Claude models will watermark the text they generate, worldwide, to comply with an EU transparency regulation. This is already live in new models released since August 2, and Anthropic says it’s coming to their other models by December, on a timeline they haven’t specified and that users can’t detect. The technique is steganography by word choice: at each point where the model picks the next word, a secret key sorts candidate words into lists, and the model leans toward one list just often enough to leave a statistical fingerprint. The words themselves are the mark; there are no hidden characters or visible markings. If this sounds a little heady and you want to see exactly how this works, James Padolsey’s interactive explainer is wonderfully clear. Playing with his examples makes the concepts click in a way simple prose can’t. Finally, only Anthropic holds the key that sorts these words into the secret lists, so only Anthropic can ever detect it (or decline to, or be wrong about it).

Three great pieces I read on this:

• John Gruber who is offended personally as a writer (though he never uses AI to write) at Anthropic’s stance that nudging word choices is imperceptible because swapping a word for a near-synonym doesn’t matter much to the reader. It’s as if someone told John that fonts don’t matter because the words are still the same (well, maybe he’s not that mad).
• RIP, Claude is Michael Lopp’s take that you can’t use Claude anymore for editing. This resonated with me, since my use case is similar.
• In Claude’s watermark is akin to an ai poison pill, Jeff Gamet pretty much says it all in the title. If this watermark could sneak in anywhere, it means that everything you use it for is suspect.

All three are great and part of why I felt like I wanted to add my perspective as someone who has recently built a writing habit with Claude.

A partner with its fingers crossed

The blank page deflates me. Give me endless open possibilities and my brain starts running a million simulations before filing the whole project under “overwhelming, do something else.” It’s a procrastination trap and I hit it hard. The way I work around it is to lower the activation energy. Find some small piece I can do right now, get the ball rolling, and then I settle into working. I love doing things, but I hate starting them. And so writing a blog post a week has always felt out of reach – I’ll waste a day staring at an empty screen, only to decide that it’s not that important and I have lots of other work to do (which I do).

What finally worked, after years of ideas existing only in conversations, was having Claude interview me. Claude lowers the activation energy. I’m answering a direct question on a topic I had an idea about and suddenly there are two thousand words of raw material in my own voice to carve into something. I’ve published a post nearly every week this year because of that process. It’s exactly the “thinking partner” use Anthropic advertises.

That entire practice is now suspect. Every post I publish now carries an invisible question mark that only Anthropic can resolve, in either direction, forever. My ideas, my answers, my editing, and their ownership. It’s like having an editor who not only requests to have their name put on the front cover, but goes so far as to list me only in the acknowledgments with a note that I didn’t contribute much and frankly wasn’t that nice to work with.

Tools, not authors

On Sharp Tech this week, Ben Thompson connected this to an argument he made four years ago. For all of human history, idea creation and idea substantiation came bundled. Having an idea wasn’t enough. To get it into the world you had to substantiate it into something – an article, an image, a film – and the work of making it into something was inextricably tied to getting that idea out in the world. AI unbundles those two things. Some of the substantiation can now be delegated, which means the unique insight that started it all can finally be valued on its own. What Anthropic and the EU are attempting is a forced re-bundling: if AI touched any part of the substantiation, then AI gets the credit for both.

Underneath that is a deeper disagreement about what AI even is. Thompson’s view, which is also mine, is that AI is a tool. Humans use tools, and humans bear the responsibility for how. When someone uses Photoshop to fake a photo of a celebrity, we blame the person. Photoshop made the fraud far easier than airbrushing a print ever was, but nobody indicts the software. Yet when someone uses AI to generate the same fake photo, suddenly we blame the AI. That ascribes moral agency to something that doesn’t have it, and it absolves the human who does. I’ve written before that AI is a fantastic tool and a terrible machine, and this is a similar thread: a tool is something a human wields and they bear the responsibility for the output.

If I generate vapid, hollow AI slop and post it as my own work, that’s my error and I have to own it. No watermark is needed to establish it; the writing convicts me on its own. And if I bring my own insight, use AI carefully as an editor, and meticulously edit every paragraph, there’s no reason to witch-hunt whether “meticulously” was suggested by AI (it wasn’t 👋) or thought of by me. Any more than there’s reason to ask whether I used a thesaurus, or to credit the spell checker for the lack of typos. These are tools, not authors.

Bundling incumbency

The other aspect of bundling is advantage and incumbency, and watermarking is an incumbent’s move: an attempt to rebuild a bottleneck that technology just removed, deciding from above what counts as legitimate creation. As Ben points out, the bundling of duplication and distribution meant newspapers not only had a solid business, but the ability to decide a lot about what ideas got disseminated. The internet removing that bottleneck has led to many negative outcomes, but it has also led to so many more talented people emerging (like John, Ben, Rands, etc.) that wouldn’t have been heard previously. It was one of the most exciting aspects of the early web and especially blogging.

The same sort of debate echoes in vibe coding and AI-assisted coding. If anyone can program, then that means people like me who can code are supposedly threatened. There’s a strong impulse to gatekeep, to loudly proclaim vibe coding as “less than”. When you’re an incumbent with a good thing going, there’s a built-in bias against tools that would allow others into “your” space.

But that’s the point of a tool – to flatten some aspect of the work so that more people are capable of it. This is precisely why tools are wonderful. More volume of things being built necessarily means an increase in the number of bad ones – software, writing, images, etc. However, it also means that the net of good things rises much faster than without the tool. I’m excited rather than threatened by AI-assisted coding because the value was never just the shape of the code and cheaper code makes more things possible. There will be (and already are) great things being built by people who are new to coding.

So whether it’s vibe coding or AI helping people write, ditch the elitist witch hunting. Judge the quality of the product instead of trying to prejudge the methods of input.