We’re in the middle of some fairly spirited debates right now, in different corners of different communities, about the use of AI in writing. And by writing, I mean the whole range of it: people using AI for LinkedIn posts, people using it to edit a manuscript, people using it to actually generate fiction from a prompt. Some people have drawn a hard line. “I’ll never read that AI crap again.” They’re looking for tools that will tell them, definitively, this was AI, so they can put the book down or skip the post. Other people have just moved forward and embraced it. And I think both sides have a point, which is exactly why the debate keeps going in circles instead of landing anywhere.
I want to walk through how I got here, because I think my own path through this is a pretty good case study for the argument I’m about to make.
How I used to do it
Years ago I published my first book on Kindle. I’ve since published two others. My workflow back then was fairly disciplined. I’d sit down and plan the series out. I’d outline it in Scrivener. Then I’d actually write, sit down and get the words on the page.
After that came the editing phase, and for me that meant a tool called ProWritingAid. At the time, it didn’t have AI built into it at all. It was rules and pattern matching. I’d run all kinds of different scans on my own writing: a spelling pass, a pass that would flag overuse of adverbs, a grammar pass, and a few others. I had this regimented thing where every chapter went through five or six of these passes. Then I’d proof the whole manuscript by hand, run one final ProWritingAid pass over the top of it, compile the manuscript in Scrivener, do the cover work, and publish.
The editor was always outside you
Go back further than that, before tools like ProWritingAid existed, and the process looked different but the shape was the same. You’d write a book and you’d pay an editor. That editor would read your manuscript and mark it up: grammar’s bad here, you’re overusing adverbs, you lost the plot line in this chapter. Whether it happened in Word or on paper, that’s what an editor did.
And the reason that role existed at all is that most writers aren’t experts in grammar, and most writers aren’t experts in the conventions of their own genre either. So you paid someone else, someone outside your own head, to catch what you couldn’t catch yourself. That’s the whole job. And when I really sit with that, I realize ProWritingAid was doing the exact same job the paid editor used to do. It just did it without a human sitting on the other end.
So look at that arc as a straight line. First, an editor, a person outside you, modifying your manuscript, and you paid for it. Then, word processing and tools like ProWritingAid, still something outside you making modifications, just automated and a lot cheaper, which is part of what made self-publishing through KDP viable for people like me in the first place. In both cases, something external to me was shaping the final text. That’s not a new idea. That’s just how writing has worked for a long time.
Where I am now
Now we’re in the AI era, and for me specifically, that means a fully agentic workflow. I dictate my book. I hand it to Claude. Claude goes through and handles the dialogue tags, the dialogue punctuation, and builds a clean markdown file out of what I gave it. And I’ve been very deliberate about the guardrails I put around that. I give it my prior books and tell it, this is my voice, learn it. Don’t invent things I didn’t say. Don’t add in superfluous wording or phrasing that’s inconsistent with how I actually write. I don’t write like AI writes by default, and I don’t want my books to start sounding like that either.
So here I am, using this in my own books, and the honest question is: will people read them? I don’t know. Maybe some people won’t, because I used AI and that’s a line for them. Other people genuinely don’t care. But I think this is only going to become more relevant over time, not less, because AI is getting baked into everything. It’s in Microsoft Word now. It’s in nearly every editor people use. It’s getting harder and harder to avoid touching it at all, even by accident.
And here’s where I land on it. To me, AI acting as an editor isn’t meaningfully different from paying a human editor to do the same job. Both are an outside viewpoint looking at my manuscript for errors I can’t see because I’m too close to it. I don’t see the real distinction. So if someone decides they’re not going to read my books because I used AI to help me get there, fine. I can live with that. But I don’t think it’s a fully coherent position anymore, and I say that as someone who’s paid for the human version of this exact service.
Editor versus author is the actual line
That said, I don’t think the debate is as flat as “AI touched it or it didn’t.” There’s a real distinction hiding inside all of this, and I think it matters more than the binary people usually argue over.
Using AI as an editor, the way I just described, is a fundamentally different thing than telling AI, write me a poem about this topic, and having it generate the whole thing from nothing. The second one is AI doing the writing for you. The first one is AI doing what an editor has always done, catching and correcting what’s already yours. Some people don’t draw that line at all. Their position is simpler and blunter: AI touched it, I’m not interested, full stop. I understand that position, but I think it collapses two very different activities into one label, and it’s going to get harder to hold as AI becomes more ambient in the tools everyone already uses.
What the watermark actually tells you
This is where the em dash debate, and the whole “AI slop” conversation, gets more interesting than people usually treat it. Anthropic recently started embedding an invisible watermark into text that Claude generates, and it’s worth understanding what that actually is, because it’s not what most people assume.
It isn’t hidden text, and it isn’t metadata stapled onto a file. It’s closer to what you’d call steganography. Every time a language model generates a word, there are usually several roughly equal candidates for what comes next. Normally a tie between them gets broken by an arbitrary random number. With the watermark, that tie gets broken instead using a secret key and the words that came just before it. The word chosen still reads completely naturally. But string enough of those word choices together, and a pattern emerges that’s statistically detectable to anyone holding the key. Nothing is added. No extra characters, no extra tokens. It’s baked into which word got picked, not bolted onto the text afterward.
Here’s the part that speaks directly to the editor versus author distinction I just made. The watermark only has something to work with where there was an actual choice to make. Correct a piece of code, answer a math problem, or fix someone else’s grammar and punctuation without touching their wording, and there usually isn’t much room for the watermark to register at all, because there’s only one right word or correction to make in each spot. The more Claude generates from a blank page, the stronger the signal. The more it’s just lightly editing something that’s already yours, the weaker, sometimes undetectably weak, that signal gets. In other words, the technology itself draws almost the exact line I’m arguing for by hand: there’s a real, measurable difference between AI writing something and AI editing something you wrote.
Even when the watermark does register, though, it can’t tell you which of those two things happened. Anthropic says this plainly: a detected watermark only tells you Claude was likely involved at some point. It cannot tell you “Claude wrote this” from “Claude heavily edited this.” So even the most precise detection method we have can’t settle the argument people are actually having.
And it’s worth separating that watermark from the other kind of “AI detection” everybody’s actually reacting to day to day, the kind based on phrasing tells. Certain constructions, certain overused words, a fondness for a particular sentence rhythm. That’s pattern matching against style, not a cryptographic signature, and it’s a completely different and far less reliable thing. It’s also exactly where the em dash argument lives. I’ve read writers who’ve always used em dashes, long before any of this, arguing they’re not going to stop now just because AI also likes them. And they shouldn’t have to. But if a reader has decided the em dash is an AI tell, that writer just got flagged as a bot for writing the way they’ve always written. That’s not a flaw in AI. That’s a flaw in using a stylistic tic as evidence of anything.
How Claude marks AI-generated content
Where I land
I don’t think there’s a clean right answer here. What I keep coming back to is that this isn’t a new kind of choice. It’s the same choice writers have been making for as long as publishing has existed: do you bring in something outside yourself to catch what you can’t see on your own, and if so, how much of the actual writing do you let it do. An editor was always outside you. ProWritingAid was always outside you. Now it’s AI, and the scope of what it can do has gotten a lot bigger, which is exactly why the debate feels so much louder this time.
Each of us gets to decide what we’re willing to read, and what we’re willing to write with. I’ve made my decision. I use AI as my editor, under strict guardrails, to protect a voice I’ve spent years building, and I don’t think that’s a different act in kind from paying someone to do the same job with a red pen. Whether you agree with me or not, I don’t think this goes away. It’s not a fad. It’s an evolution of the tools writers have always reached for, and the argument over where to draw the line is going to keep going precisely because there isn’t one clean place to draw it.