As someone with a heavy interest in both creative writing and artificial intelligence, I often get asked “do you use AI to write?”

And the answer is … well, not really, but only in the sense that I’m not having the LLMs write words for me. There are lots of ways to interface with an LLM, and I do interface with them as part of my workflow, but the LLM isn’t “doing the writing”, at least in my opinion. Let’s get right to the use cases.

Research

I was recently writing a scene where a character was describing a procedure done on a finger without anesthesia. She called it one of the most painful things a person could experience. Then I thought to myself “wait, is that true?” then “how do you amputate a finger anyway?”

So I went to my default, Claude Opus 4.5 with long-term memory turned off and asked:

What is involved in a partial amputation of a finger?

So Claude spits back some information, which I trust to be generally accurate as I’m asking about a general subject that it should have a lot of information on. There’s stuff about trimming down the bone that makes my skin crawl, and soft tissue flaps, and all kinds of things, and I can pretty easily drill down to ask about specific procedures. Best of all, I’m being given keywords to search with, like a “rongeur”, which are cupped pliers used on the bone: nicely evocative for my writing. This is also something I can search up and get an image for.

My previous research method would be to go to Wikipedia and see whether there was a relevant page, or just Google the question directly to see whether I could find some simple guide, and failing that, ask someone I know with relevant expertise (medical, in this case). I still do these things sometimes, but for a purpose like this, the LLM is very fast, and the results tend to be easy to verify. Wikipedia does have a page on amputation, but nothing on finger amputation generally. One of the pitfalls of research as a writer is being unable to find generalist sources while being too out-of-depth to easily understand the specialist sources, and the LLMs can synthesize for you a nice middle ground. Wikipedia is, of course, still my go-to, especially after drilling down with Claude a bit to get some keywords and a broad zero-context overview.

There are, in my opinion, two major risks with using an LLM for writing research.

First, the LLM will bullshit sometimes. The models have gotten better about this than they were two years ago, but it does still happen, and happens more often when you’re probing a specific area of expertise where it just doesn’t have enough data. The models have a desire to please, and Opus 4.5, at least, will also search the internet if it thinks that it’s out of its depth, but sometimes it will have something like foggy memories and not be totally accurate. I’m in the habit of double-checking what I’m told, but usually it’s easier to double-check with the proper suite of keywords and concepts than it is to go in and do research cold. Because I’m writing fiction, and often fantasy or scifi, I have a lot more leeway in getting things wrong, or “not quite right”: I think I would kick up my diligence quite a bit if I were writing non-fiction.

Second, the LLM is not good at niche stuff, and when you’re writing, you want the niche stuff. If you want to know what clothes people wore in 1890s Spain, the LLM cannot be depended upon, not unless you’re willing to be sloppy … and using the LLM encourages you to be sloppy. So if I find myself needing expert guidance on something, especially if it will be crucial to the plot, I avoid using the LLMs. The LLMs can still be used as a research tool this way by scouring the web for you and returning relevant websites, but then you’re just using a search engine secondhand (one that cannot always separate out authoritative sources from slop, not that regular search engines can).

SPAG Check

SPAG is short for “spelling and grammar” and if prompted with “check this for SPAG” the LLMs will generally do an okay job pointing out any errors. In fact, most spellcheckers have incorporated neural nets for quite a while, rather than going by strict rules or dictionaries, so this isn’t too far from what’s already happening if you have those turned on, it’s just that those tend to be specialist systems.

I consider this sort of editing to be above reproach, because the LLM is checking for mechanical errors and pointing them out to you, rather than directly editing the text. If I write:

I set me phone down on the table.

Then this is obviously incorrect and it’s just a matter of me actually being able to see this during an editing pass … or I can have the LLM point it out to me and save editing energy for other issues I want to address when I do the editing pass. Generally speaking, I think I would catch something this egregious, but I don’t always, and if I’m attempting to write 10K words a week, there’s plenty of room for error.

There are occasional hallucinations, especially if the text is clean: the LLM wants to find errors. But by acting as the intermediary between the text and the LLM, there’s no risk here, which is why one of my rules is to never let an LLM touch the text itself, nor to allow any automated process any control of the text. If there’s an error to fix, I’ll look at the section the LLM is reporting in one window while I find and address the issue on my own. I want full control over what’s happening, and a full understanding of why it’s happening, and so long as I have those two things, I’m golden.

The main thing here is that this is mechanistic editing, those parts of writing where there is something unambiguously wrong, where any writer will immediately say “oh, right, duh”. As soon as you get into questions of style, like the use of passive voice or how many adjectives to use, the feedback gets much less useful, because it depends on the subjective tastes of the LLM.

I ran this post through Claude for SPAG, and it told me that “laying” is the transitive verb while “lying” is the intransitive verb, a rule that I never remember, and which I wouldn’t have caught myself on my own editing pass.

Zero Calorie Feedback

Sometimes, in the course of writing, you need feedback. Ideally, this would come from a person you’re very close to who gets you and your writing, and also has a good critical eye and understanding of writing craft, as well as a very good model of your target audience. They also respond to requests within seconds.

In practice, this is a brutally difficult thing to find. Sometimes you ship something off to a friend who has agreed to read over your work and they’re great, but they take three days to respond. Sometimes you don’t fully align with the beta reader. Sometimes the feedback comes from a place of wanting to move the story toward being something totally different. I’ve had some unhelpful experiences, and with a few exceptions, just don’t use beta readers at all.

But sometimes you get that itch to share a thing with someone, just to get a gut check on whether it’s working, to know whether the vibes are off. You want to bring in someone who can see the forest for the trees. And for this … sometimes I think an LLM can be okay?

My preferred method of doing this is to throw a piece of text into an LLM and just say something like “Read this and tell me what you think.” With that instruction, Claude Opus 4.5 will say something about the story, what it thinks of the prose and what works well, as well as some points it thinks are weak or need strengthening. This is very minimal prompting, and it’s minimal because I can be sure that I’m getting a very generic response; ideally, the response is a similar response to what I would get from the generic reader.

Will the LLM be correct? Well … no. Or at least, my experience has been that I’ll only agree with it on occasion. But it does give a little bump of “alright ‘someone’ has read this” and that can be heartening and help along the writing process. Sometimes when writing you get so into the weeds that it’s hard to know what’s going on, what’s working, whether it’s shit or not. The LLM says some things, and sometimes that works for me.

In terms of what’s actionable? Eh. I fed a chapter in, and it said that the tour section went a little long, and this did make me reread that section, and I did agree that this slowed down the pacing in a place where it didn’t need to be slowed down. In another place, it criticized a description as lingering too long, and I frowned at the passage before deciding that it was fine. The experience of doing this has been mixed, but I think on the whole it’s often worth it, particularly when a chapter is early in the editing process.

I haven’t found sycophancy to be too much of a problem, but it’s very prompt-dependent. If you say that you’re an editor at a publication house and you want to know whether a piece deserves to be drawn out of the slush pile, it’ll be much more critical. If you present a work as your own piece, it’ll be much more kind. But I don’t think you generally want a maximally critical approach, because the LLM is attempting to give you what you want, and it’ll be more prone to seeing problems that aren’t there. My opinion is that neutral evaluation is likely to see things that are “real” at least some of the time, and it won’t often make up descriptions of the work that are entirely off. I do try to prompt in such a way that the LLM does not assume that this is my own writing, and this does help to bring sycophancy down significantly; we are, together, saying what we think of this piece that we found lying on the ground somewhere.

Is a human reader better? Yes, obviously, if they’re a good human reader with good feedback who will get to the piece in a timely manner. But I think most writers have at least some experience with how frustrating it can be to get eyes on something they’ve written even after it’s been polished.

Brainstorming and Idea Bashing

In computer programming we have a concept of “rubber duck debugging” where you have a small rubber duck (or other non-sentient object) that sits on your desk for you to talk to in order to work through a problem. This helps by requiring you to put yourself in a different, outsider mindset, and switches up how your brain works. I carry around an imaginary rubber duck all the time, and use this technique a lot.

Brainstorming with an LLM is not the same as brainstorming with a rubber duck. Instead, brainstorming with an LLM is like brainstorming with a guy who doesn’t have a clue what you’re talking about, and has been cursed by a genie to only give bad, rote ideas.

If you’re thinking that doesn’t sound helpful … well, it’s an acquired taste, and one that I haven’t particularly acquired. The only times it’s really worked for me are when I frown sadly at the screen and then start typing out “that’s not how you do it, this is how you do it”, which can help to get the juices flowing. But then again, I’ve always found the ideas to come pretty easily, and finding a seam of good meat has never been too hard.

LLM ideas tend to be bad because they’re obvious, and the last thing that you want a reader to think when reading is “well that’s obvious”. There are tricks you can use to try to get the LLMs to have better ideas, but none of the ones I’ve tried have had particularly good outcomes, places where I’ve thought “wow, that’s brilliant”. In the current generation of LLMs, at least, “wow, that’s brilliant” is not something that you’ll hear a lot. Instead, you’ll hear “that would have taken me a long time”.

I don’t think it’s entirely worthless to talk through ideas with an LLM, but it’s like talking to a friend with bad taste who Doesn’t Get It. People have told me that this can be ameliorated by pushing more things into context and “properly prompting” and that hasn’t been my experience, but I’m picky. What I want from brainstorming is a seed of an idea that latches onto my mind, and I’ve never gotten that. Your mileage might vary.

Prose Generation

I don’t use LLMs for prose generation. I think they’re not good writers, generally speaking. I’ve tested this pretty extensively, and there’s a tendency for what they write to come out safe, flat, and unmotivated. Sometimes I’ll have a good idea, write a short story, and then give the idea to the LLM to see how it iterates on and executes that idea, and I’ve never preferred the LLM’s version, even with prompting to push it more toward my preferences. There are occasionally nice lines, in isolation, but they’re not put together well, and get repetitive quickly.

So this is a pass from me, personally.

In theory you could have an LLM produce a bunch of text and then edit it into shape, but I’m not convinced that this is actually less work than just writing the thing. I told myself I was going to attempt it once, and lost all enthusiasm for it after I started reading the story.

There’s another reason not to do this: people hate it. LLM-generated text is so widely known to be bad that admitting you started from an LLM-generated base is going to kill any desire to read it. This is aside from the emotional response that people have to artificial intelligence, which, at the time of this writing, I would also expect to be bad.

And yes, I think you do have an ethical duty to inform readers if the text was AI-generated, in whole or in part. Because people have stated strong preferences, you have to report. I personally don’t have any interest in reading the dull prose of an LLM, and if it’s a factual post, I would rather just query the LLM myself, if that’s the level of insight and quality I want. “This was AI-written” is a strong signal not to read, unfortunately.

I think a year ago I’d have told you that I expected either one of the frontier labs or an intrepid startup to come along and change things in the near future, building some kind of model or harness on a model that could bang out a stellar novel that was well worth reading. I’m more skeptical about this happening now, for all that I don’t think there’s anything special about being human or having a soul to your writing. There are definitely people peddling AI writing and AI-cowritten stuff right now, most of them doing it stealthily, and I’m sure that at least some of it is making money … but I haven’t read anything that wowed me, at least not yet.

Now if you’re a bad writer, the LLM might be better than you already. If that’s the case, I can see the temptation, but it does mean that you’re not going to get any better at writing, and you probably won’t be able to diagnose where the LLM is going wrong. But perhaps this gives some outlet to ideas that have been floating around in your head somehow. I’m less sure about this, because it’s not the position that I’m in.

I’ve talked to at least one author who used an LLM in their workflow, but it was mostly as a method of getting unstuck and staying in flow state. I’m not sure exactly how it worked for them, or what percent of the text was “LLM generated” in some sense.

(Personally, my prose is me, and the sort of thing that I wouldn’t be willing to give up even if the LLMs could do it better. And the fact that it’s me is also my comparative advantage when it comes to getting people to read my stuff.)

Will It Get Better?

Probably, yeah. Maybe the rumored next models from OpenAI and Anthropic will just totally knock it out of the park when it comes to writing. LLMs are still evolving, and we’re learning to use them better, putting harnesses and prompting structures in place. It’s possible that someone will come up with a better multi-agent system for creative writing that allows it to be a better assistant in various ways, less prone to failure or bad feedback.

That said, I don’t expect my own use to evolve, and if we ever get to the point where I think any of the LLMs are a better writer than I am, I think I’d still silo myself in my own home of words, rather than giving up creative control.