Last time, I covered what ChatGPT is, what it isn’t, and some things to keep in mind if you choose to use it to help your fiction. Now we’re going to get into how to use it, or other machine learning programs, to aid your fiction. Though, first, I’m going to try to underscore some of the caution I tried to instill in the last post: do not mistake ChatGPT for an unbiased assistant, talking encyclopedia, or genius author.

ChatGPT is Your Tool, Not Your Coauthor

As I said last time, ChatGPT is a particular tool. On his WriterDojo podcast, Larry Correia frequently describes elements of writing as “another tool in your toolbox,” meaning that you don’t have one tool for all jobs, and not all jobs require all of your tools. It also means you should put in the effort to understand the contents of your toolbox; you can technically split a log with a hammer, but that doesn’t make it a saw.

Similarly, ChatGPT is one tool. It can’t be used for all things, it can’t write your book for you, and possessing the tool does not impart mystical knowledge on how to best use it. When you use ChatGPT for anything, you have to know enough about the subject to be able to do it yourself. You need to know enough to figure out when it’s wrong, which can be often. As I said in Part One, it’s not actually an AI; it’s a language model that predicts the next word in a sentence, and the next sentence in a paragraph, and so on. It can’t reason, it can’t understand drama, and it has no actual creativity. It runs on statistics.

What it is, is a force multiplier. Just because you can do something doesn’t mean you should spend all your time doing it. Just because you can write a book with a quill pen doesn’t mean it’s not faster to use a typewriter. I know for certain that some of you write on paper at least part of the time, because I know some of my readers from writer groups I’m in, and it works for you. Use the right tool for the task at hand.

Your Tool is Only as Useful as You Are

In a nutshell, using ChatGPT for writing is a faster version of combing through multiple articles looking for research, character elements, worldbuilding details, and so on. One of the examples we’ll look at in the future is using ChatGPT to build up a medieval village for a fantasy setting, and it’s a lot faster than reading through a half-dozen articles and a couple of books on real-world medieval villages to get things right. That’s because you’re not writing an academic paper, where you have to be accurate; you’re writing fiction, where you have to be convincing. There’s a difference, and while ChatGPT can’t be trusted for accuracy, it can be used to jumpstart convincing details.

But, again, you need to know enough to know what suggestions work best and which don’t. There are plenty of details authors get wrong all the time, only some of which can be fudged by “It’s not Earth history” or “I want to do things differently.” However, most of them do not matter because your audience isn’t going to notice. You’re aiming for less than fifteen percent of your audience noticing the errors, and that’s part of why people hire developmental editors like me to make that percentage as small as possible; but you’re going to hit a point of diminishing returns no matter how accurate you are. So you don’t have to be an expert; just knowledgeable enough to be convincing.

Also, because ChatGPT statistically models responses based on general data, that means that when it comes to modeling general opinions, its opinion is only as good as the average human who has written something on the Internet about the subject. That alone should give people pause when treating ChatGPT as infallible. I mean, have you seen the Internet?

All that being said, let’s look at an overview of ways to use ChatGPT. I will probably do dedicated blog posts on each of these topics, but for now here are some ways to use the program.

Outlining and Plot Twists

Let’s start with the least useful ways to use ChatGPT. As I already said, but it’s worth repeating, the computer has no clue how to handle drama. The outlines suggested by ChatGPT are either too generic or go in strange directions; remember, it’s based on a language model that uses statistical extrapolation, so if you feed it generic stuff it’s going to give generic responses. The more you specific you make your prompt, the more unique the answer; but with a topic as large as a novel you’re probably going to put so much work into refining the outline that you might as well have started from scratch anyway.

Plot twists are hit-or-miss, but slightly more hit than outlines. Again, the program chooses the most statistically likely next result based on prior information, so that can be both good and bad for plot twists. A good twist should feel inevitable, but only in hindsight after the audience is surprised. You can use this to your advantage when you know how to work backward to properly disguise the obvious outcome, but if you’re already good at that you probably don’t need ChatGPT’s help. I suspect that its best use when it comes to plot twists is to generate writing prompts for creative writing teachers to give to students to “fix.”

That said, you can use it to make suggestions about events. Again, you’re probably better off doing this yourself, but ChatGPT is good at doing if/then prompts, especially for things that happen off-screen; for example, you might need to show that the kingdom suffered a devastating setback in the initial battles of a war, so that your baker’s son character can see the soldiers’ faces go from optimism to despair and hear the stories of what happened; but what, exactly, was the big setback? Well, you could get ChatGPT to make suggestions — and I do mean suggestions, plural, because being able to pick and choose (or mix and match) is helpful.

The most useful case in this category, however, is asking it to list the tropes and elements of a particular sub-genre. This is useful because its statistical model appears to generate a consensus from across the Internet, which makes it a rare case where it actually can be unbiased, at least compared to any one human. It makes it difficult to generate, say, an unbiased consensus on a political topic, but for creative writing it’s actually useful. This can let you use ChatGPT to check what you’re thinking against a list of the most common things your audience may expect from a particular genre, and so avoid tunnel vision. It’s more common than most people know for an author to think they were writing Genre A only for their editor to point out it’s really Genre B with a dash of C. Even when the author knows what genres they’re playing around with (for example, my first editing job was entirely due to the chief editor reading my then-current work in progress that was Genre A deliberately dressed up in the tropes of Genre B), it’s important to review the tropes you’re working with. Genre is primarily a marketing tool rather than writing per se, but it’s helpful to keep in mind so you don’t make accidental promises to the reader that you won’t be able to deliver.

Dialog, Characterization, and Motivation

First, you have to understand that ChatGPT is no better at writing scenes than it is at writing novel outlines. However, its statistical model can be helpful when writing dialog because it’s specifically trained to analyze the likelihood of a particular response to a particular statement. The execution will always sound like a computer, and the computer will have less understanding of drama than a typical nine-year-old . . . but if you happen to be struggling with dialog, I can imagine it’s a useful tool as long as you completely rewrite it. I don’t happen to struggle with dialog, but then advice on natural dialog is one of the things authors like to get from me, so if it’s not your strong point then this might help.

Personally, the main way I think I’d use it to generate dialog is for incidental speech, such as the classic “character overhears snippets of conversation in a crowded room” scenario. Being able to quickly generate an outline of twelve different conversations from simple prompts can be a time-saver. Even when all you want is a line or two from each, you still have to come up with plausible topics. This is a niche situation, but any time I need to generate it on my own I find it irritating because it can be a lot of work for little return. If you’re writing, say, Regency-style socialites, this could be very useful.

Characterization and motivation are similar situations; if you’re having a lot of trouble with those, you’re having trouble with writing in general. However, there come many moments where you sit back and think “Okay, why should the character follow the plot?” You’ve had it all set in your head for months, and now you’ve gotten to the scene and realized something is missing. Being able to plug in a prompt like Please generate five suggestions on why John would slip out of the party early can help with those moments when you got stumped. It happens more frequently than we’d like to admit.

Idea Winnowing

You’ll likely see ChatGPT recommended as a tool for “idea generation.” This is a waste of your time. Ideas are a dime a dozen and your head is already full of more than you can use. If an author needs help with ideas, it’s in winnowing those ideas down to something useable. Many authors, even experienced ones, have to deal with something called decision lock, where you can’t pick between two or more choices and start going through them over and over, second-guessing yourself and wasting valuable writing time.

Instead, you can feed certain prompts into ChatGPT, like Please generate a background for this character using these parameters. I’ll be doing a dedicated post on this in the future and how to do follow-up prompts, but as a quick example, let’s say you want to create a secondary character who’s an apprentice wizard. You might use a prompt detailing her age, region of origin, social class, and basic personality and then ask fora physical description, suggestions on her hobibies, her magical specialities, her home life, and what topics she likes to read about besides magic. Or perhaps you want a quick description of the goods being carried by a merchant, or maybe the typical office cubical clutter of a programmer who’s only in one scene.

If you can’t come up with those yourself or know how to look up the details, then you have bigger problems with your writing than just getting stumped. However, if you’re suffering from decision lock, blanking on what the typical trade goods of medieval Germany might have been, or it’s going to take a while to get a response from your computer programmer friend who’s busy at work doing actual programming, ChatGPT can speed up the process and let you get back to writing faster. Even if you’re taking my advice and worrying about that on a later draft, anything that speeds you up without compromising your quality is worth considering.

You do still need to have passing familiarity with the topic at hand, of course. ChatGPT isn’t going to necessarily catch anachronisms if, for example, your programmer scene is part of a Cold War drama set in 1984 and ChatGPT suggests a book about C++ is on the desk. C++ was released in 1985, so you just gave anyone who knows that incentive to write you a bad review, and “ChatGPT lied to me!” isn’t going to be an acceptable excuse. At the very least, you’re not going to be able to replace alpha and beta readers with machine learning.

Worldbuilding

This is, in my opinion, the place where ChatGPT shines as a writer’s tool. Again, the vast majority of your worldbuilding will be done by you; but this is a great tool for filling in the blanks, generating details fast, and prompting yourself to think of filling in aspects you might not normally have come up with so early.

Sometimes the details in question are things that a good worldbuilder knows to leave for later. What are the imports and exports of a character’s hometown off in the frontier where orc raids are common and roads are not? Who really cares unless it’s relevant to the plot, right? Well, often those details only get filled in when they do become relevant because it’s not worth the effort until you know you need it. Instead, in the space of half an hour, I was able to go through several prompts, many of which only occurred to me after reading ChatGPT’s responses, and generate a detailed overview of the home region of a character for a book I want to write. At least a fourth of what the computer generated was useless, but that didn’t cost me anything other than the time it took to read it and reject it. The computer even suggested things I didn’t think of, such as which fish species were in the river.

The best way to do this is to suggest a real-world location to use; for the above-mentioned region, I used the Scottish Highlands as a model. Other experiments I’ve done have included Eastern and Central Europe, the American Southwest, and getting it to reinvent Skyrim (The Elder Scrolls franchise), Stormwind (from the Warcraft franchise), and the titular space station from Babylon 5. As long as it’s a place loaded into the database, it can work. I’ve even used two completely different regions at the same time (Scottish Highlands, but with the climate, latitude, and biome of the American Southwest), but that’s more finicky because the statistical model isn’t always sure which way to jump.

Very frequently, though, you’ll find yourself thinking “That’s crap, but it makes me think of something better.” Much of using ChatGPT is as a virtual sounding board, so you can bounce out ideas and see what it looks like when phrased by someone other than the voices in your own head.

The next few installments in this series will look at specific examples from ChatGPT. Will I run out of robot/ChatGPT-related memes? Stay tuned!