I use AI in my writing. I don’t think this is controversial - not necessarily - although I think it’s important to acknowledge and address it, because for some authors, it can be and should be controversial.
I don’t mind using AI. I do have a problem with AI ghostwriting content. I think that’s where the line is, and it’s a sliding scale when it comes to deciding specifics.
How I Use AI
I use AI as a soundboard, primarily. Here’s how that works:
Seed Idea
I come up with an idea, usually expressed as a thesis statement - although it can be a thesis idea, I guess, and that’s the seed that might create content.
If I feel like the idea has enough clarity - i.e., I know how I’d like to address that idea being written - I’ll crank up an editor of some kind (usually Obsidian, but might be Word, or Scrivener, Sublime Text, or even Joe's Own Editor, and this isn’t the full list of options!) and just… flow. I’ll write whatever I think of, in the form that I think of it, trying to form a general narrative.
It’s usually formatted in MarkDown because it’s stupid-simple to write and requires little cognitive space. I want to emphasize something? I use asterisks, and most other formatting things flow pretty naturally in how I’d type on IRC or email or, well, pretty much anything.
| For the record, the editor I’m typing in right now - as I write the first draft - is Sublime Text… although I chose AsciiDoc for the format on a whim. I’m whimsical that way. |
But what if I feel like I don’t have enough clarity? What if I think the seed idea is worth exploring, but I don’t really sense a direction for it, as the seed thought manifests itself?
Well, I used to wonder if I had any fellow authors who got me - and the domains in which I write - well enough that I could bounce ideas around. But that takes their time, and relies on them knowing how I think (which means it’s a small set of people in the first place) and how I write (which means the set is even smaller). That makes bouncing ideas around feel like an imposition, and I don’t like imposing on people, even if they’re willing to endure the imposition.
Enter AI: I can craft a query about that seed thought, asking for early impressions or what other people have written about. I tend to write that initial query in such a way that I’m discouraging overly positive feedback - because the AIs sometimes tend to respond as if you’re simply The Most Special Boy Who Ever Specialed - and I’m interested usually in whether the idea has legs, or if it’s actually a super-common thought that everybody has, and that’s why nobody has written about it ("Article: The Sky… is BLUE!!!!").
I might also ask for a possible direction the idea could take, if I’m struggling to establish a narrative.
I tend to mutate any suggested narratives pretty strongly. I’ve certainly used them, but the ideas I have about narrative tend to not mesh extraordinarily well with the ideas the LLMs' models have about narratives, plus the LLMs have a tendency to emphasize positive reinforcement and positive outcomes more than I do.
So: first, the AIs are soundboards for core ideas.
Refinement
After I’ve used the AI as a soundboard to refine (or kickstart) a draft, well, I write the draft. I usually write in shifts, so I’ll spend a while writing that draft - and doing little else - and at the end I’ll have a few hundred to a few thousand words.
I’ll do a manual review of the draft until I think I’d be okay publishing it as is. Interactions with any AIs here are limited, mostly to Grammarly, an AI tool that assists with grammar and phrasing - and honestly, that’s more likely to address typing errors over anything else. I’m writing in my voice at this point, including any oddities in how I describe things. (This usually means lots of dashes, commas, parentheticals, and semicolons, and some weird wording and colloquialisms; I write weird, y’all.)
Once I have the draft down to where I would be willing to publish it myself, shrugging at any problems that are manifest in how I write innately, it’s time for the AIs to get back in the ring for real.
At this point, I’ll copy/paste the draft content for the AIs. I’ll pick the most advanced models available to me, and ask for critique for style, effectiveness, and accuracy, and please, avoid unnecessary positive reinforcement. (This is not my actual query, but this is pretty much what I’m asking from the models.) I’ll include the title, tags, any potential imagery ideas, and ask for things I’ve omitted, as well as suggestions for a better title, or tags, or… well, you get the idea, I hope.
I’m asking for the AI to check my work, to steelman it, to make it stronger. What did I get wrong, what could I do better?
This is usually the only stage where AI contributes more than feedback - though still only as a critic, not a ghostwriter. Usually the AI complains about my colloquialisms and syntax, and it wants me to add headings because my draft flow is often (quite) uneven. (For the record, this entry got the 'uneven flow' criticism, with the middle sections being "too long" compared to the other sections. C’est la vie.) More bullet points, and often far tighter text; the AIs don’t care for narration much.
For example, the first draft had This is where the AI "writes." ChatGPT suggested This is the only stage where AI contributes more than feedback—though still only as a critic, not a ghostwriter., and by golly, I like that better. And I used it, although I also corrected the use of the dash to fit how I format dashes.
|
I ignore a lot of those suggestions… but not all of them. Often I’ll go back and blend what the AI suggests in, occasionally verbatim but usually rewritten in my own voicing. Occasionally I’ll change the flow to fit what the AI recommends, and I have been trying to use headings more than I used to.
If the AI says "You’re wrong here, and here’s why," well, that’s a pretty impactful critique, and I take those notes very seriously. If I have to rewrite sections - or scrap the entire post because it’s built on an incorrect assertion - so be it, unless I’m going to rewrite the piece to address why I thought the wrong thing.
I’ll keep going back to the LLMs until every criticism it has is looping back on the same concerns I’ve decided I don’t care about, those comments regarding voice or possible reader critiques that I’ve decided are worth accepting.
How People Use AI Badly
I’m still a little conflicted in how I use AI - I’ll avoid copy/paste in general, especially for stretches longer than a sentence or so, but that’s an evolving process.
What bugs me about AI is when people use it for seed ideas or writing complete drafts. I’ve tried that, just to see if it worked, and honestly, it really doesn’t. The AIs tend to create content that feels vaguely like a sort of shadow template ("Content about that should look like this") with a sort of bland unsatisfying conclusion that feels like every other conclusion out there. The AI wants to feel definite about certain things, and that just doesn’t actually resonate with me very well.
The linguistics AI uses are also pretty distinctive, and that truly annoys me, partly because I find I can lean into some of those same patterns without meaning to! For example, the use of dashes is a good marker for AI content - particularly the em-dash, a "dash the width of the letter m." Grammarly will suggest an em-dash on occasion, and I use some editors that like the em-dash, too. That’s a concern because it affects how people read content ("Oh look, an em-dash! Was this generated by an AI?") and I don’t want my readers predisposed if I can help it.
I feel like when people rely on the AI to write content too much it feels very milquetoast, very generic. It has a sameness to it I find unappealing, and it’s horribly lazy.
Being a lazy writer is unfortunate. It’d be better to be silent than to be a lazy writer.
How I Feel About The AI
I had a three-paragraph meandering about hammers and carpenters and fences - and ChatGPT zeroed in on the concept quite well, and it suggested something like this content, which I found more effective than my draft:
Using AI in writing should be like using a hammer to build a fence. You still have to measure, cut, and choose the wood. But if all you’re doing is snapping in prefab panels, can you really say you built it?
That’s where I am with how I use AI; I use it as a soundboard, not a replacement for my own thoughts and expressions, and I’m fine with anyone else who uses it in the same way. And for those who go further, well, it depends on how well they blend themselves with the machine: if it’s still them, well, I’m all right - but I find listening to the machine unfortunate.