The problem that Large Language Model A.I. poses for book publishing isn’t really a new one, but one that is as old as publishing itself. A lot of people want to get their books published even when the books just aren’t good.

Early in the history of publishing, this fact meant that publishers had to wade through piles of unsolicited submissions–the “slush pile”–to find the few good manuscripts they’d received. Then word processing made typing so much easier that anyone with an idea could type up a novel, and publishers effectively handed off some of the slush by insisting on agented submissions only. More recently, when Amazon made self-publishing easy and free, Kindle Unlimited became a slush pile. Now, with the assistance of A.I., that pile is growing at an incredible rate.

I’m not going to address the problem of dealing with slush. Maybe A.I. can be trained to weed out some forms of bad writing by looking for proxies of it–but I suspect that’s a “build a better mousetrap” kind of problem. Rather, I want to ask whether A.I. can play a useful role in writing good fiction.

So what are the current strengths and weaknesses of A.I.?

A.I. can generate large amounts of text, relatively quickly, based on the writing that it was trained on. Given an appropriate prompt and context, it can identify popular tropes of various genres and incorporate them into its writing, along with some random elements.* Apparently, it can also maintain continuity in its writing, which amazes me. However, A.I. cannot respond to writing as a human does, because it isn’t human. It cannot evaluate a piece of writing and say “This is intriguing!” or “I love the way this is written,” or “I was rolling on the floor laughing.”

I think the inability to respond in a human way to a piece of writing is crucial, and this inability creates a further limitation that I’d like to point out. A.I. cannot adjust its writing to changing times (and times are always changing) except through a diet of new, human-written material expressing the varied human responses to those changes. For example, if self-driven cars become ubiquitous, people will have to write about their experiences with them, and AI will have to be trained on that writing, for it to be able to incorporate that change.

Now let’s look at an example of a writer’s process. She starts with an idea–she’ll combine some tropes popular with her young daughter and wrote a story about a girl who is secretly a princess (but doesn’t know it), who takes care of her little brother and sister when their parents are lost/kidnapped/killed, who is accompanied by a large, friendly predator (say, a white wolf with green eyes,) and who has several items of magic jewelry. But that isn’t enough. To have a story, something’s got to happen. Our writer decides the heroine must rescue her parents (they were kidnapped) from the clutches of an evil king/wizard.

Now the writer needs to work out some backstory. Why doesn’t the girl know she’s a princess? How did she come to have a wolf-friend? Why did the villain take her parents, and how?

If the author is a pantser, she starts writing and answers these questions as she goes. New ideas come to her as she types. Maybe one of the magic items is a necklace that can light up the dark forest or a dark tunnel as the children travel in search of their parents. Maybe the parents were kidnapped because, being royal, they know some secret that only the royal family knows. Maybe there is both an evil king and an evil wizard, and since they don’t entirely trust each other, our heroine can set them against each other.

If the author is a plotter, on the other hand, she will work out most of these details before she spends a lot of time typing. Once she knows with sufficient detail what’s meant to happen, she’ll set about writing it in an interesting way.

Either way, the result is a rough draft. Now it’s time to revise. The author is also a reader of books, and she uses this to evaluate what she’s written. How can it be made more interesting? Does it drag in the middle? Is the ending cliched? How can she end it differently?

She changes her word-choices, making paragraphs and sentences sharper, or funnier, or more dramatic. Again, she measures this by her own, human reaction to them. She also checks for mistakes and problems in continuity. How many horses would you typically need to pull a carriage? Could the peddler have made it all the way to the place they meet him on foot? Are there conflicting explanations of the queen’s background?

Now she has a second draft. But she isn’t the only target audience for her story. So she persuades some other humans to read it and evaluate it, pointing out problems she has missed and noting where the story fails to satisfy them as readers. Then she revises again.

At some point, the author (and her friends) decide that she has done all she can to improve the story, and she starts the process of querying agents. If she does get an agent, there will be yet more revisions in store, and so on, until the story is either published or set aside.

So what parts of this process could A.I. assist with?

First, it could generate plot ideas using tropes and random elements suitable to the desired genre. However, it wouldn’t be able to evaluate these ideas as a human would. (E.g. “That sounds intriguing!” or “I’d like to read that one.”)

So let’s say the human author steps in here, choosing from among the generated ideas the one that most provokes her interest, and then modifying it to make it even more interesting.

Speaking for myself, I don’t have trouble coming up with ideas. However, I do sometimes have trouble coming up with a suitable combination of ideas. For example, after reading an number of fantasy books in which young women living in evil kingdoms are trained to fight and kill, I decided to write a fantasy in which there was a academy where young people were trained to interact non-violently. But I still needed a problem to be solved. A violent foe? Trouble from within the school? Maybe something in the heroine’s personal life? There are lots of ways of generating ideas (roll story dice, ask a friend.) A.I. could be yet another one.

Once we have the idea for the plot, we need to develop the characters, the setting, and the backstory. Can A.I. generate backstory to explain how things came to be the way they are, and why the characters are in their current situations, and do so in a way that serves the plot? I’m skeptical, though I suppose it could suggest some possibilities. The author would still have to decide what worked for the story they want to tell

Now we get to the actual writing. A.I. can certainly generate a lot of text, but I don’t believe it can evaluate that text as it goes to see if the story is developing in an interesting way. The author can do this, BUT–important point–it takes time to read versions of a story. The would-be author could have A.I. churn out multiple versions of a story based on the chosen plot, but then she’d have to read all those versions to see if any of them were good enough to take further. If she could find a good story after reading just ten versions, that would be feasible. If, on the other hand, there was only one really promising story out of every one hundred, that would be inefficient.

That’s the reason agents tend to look at only the first ten pages of a submission, not the whole thing. From those ten pages, agents are deciding whether the author is good at writing and also good at identifying what will interest the reader. From that, plus the brief description in the query letter, the agent can judge whether this particular story seems promising. A good author with a good idea makes it worth reading beyond 10 pages. Otherwise, better to stop and move on. But this shortcut depends on the human author being deeply involved in the writing of the whole submission.

The more the author guides the A.I., the sooner she will get a promising draft. So perhaps she goes chapter by chapter, adding further details to her prompt and more context each time, fleshing out the plot in a way she finds pleasing as a human reader. (This process would probably be easier for a plotter than a pantser.)

Notice that this process is becoming more and more like writing the darn thing herself, except that at this point the writing style is the A.I.’s. Speaking for myself, getting the details sorted out is the hardest part of writing fiction. Writing the actual words is easy. (Though polishing them is another story.)

Finally the author has a promising rough draft. But she isn’t done yet. To have a really good book, she’s going to have to revise it.

Here I think the A.I. will not be very useful. Revising is all about changing the draft in accordance with her human responses to the writing. The A.I. can proofread the spelling and grammar. Maybe it can even apply some rules such as the rule of three, but it can’t respond to the writing as a human would.

In the process of revising the novel, the author’s word choices will start to replace the A.I.’s. The style will become increasingly the author’s own. (Hopefully for the better.) Since I’m assuming an author who is trying to write a good book, by the time she is done, she will have gone over every sentence, either approving it or changing it. It will be largely her own writing.

To summarize: how can the A.I. assist the author? First, it can generate ideas, perhaps ideas based on context that the author provides it, from which the author can take inspiration. I don’t just mean at the beginning–presumably the author could ask it partway through for suggestions on what might happen next.

Second, apparently it can help flag continuity problems, though I don’t have personal experience with this. Honestly, I find this the most exciting of the possible uses. Trying to keep straight everything I’ve said so far about a world, even if I try to make lists of characters, places, etc., is one of the things that gives me a lot of trouble.

Third, it can check spelling and grammar, as well as flag words that may be overused and other stylistic details.

I have assumed in this post that A.I. won’t eventually become capable of predicting (though not feeling) human emotional reactions to text. Given all the developments we’ve seen, I don’t think I can say confidently that this will never happen, but it certainly isn’t the case now.

A.I. isn’t the problem for book publishing. Authors with low or no standards are the problem, and A.I. is making this problem so, so much worse. But for authors who want to write the best books they can, A.I. could become a useful tool, along with the word processor, writing software such as Scrivener, and search engines like Google.

Till next post.

*For a simplified example of such a prompt, “Give me a 2,000 word science-fiction short story about a group of people who live 800+ years (‘the Ancients’) and how studying them provokes discord and eventually ends in violence. Set it in the US and include a protest group, a pharmaceutical company, and a cell biologist.” This is based on a story that I am working on (not with A.I. assistance.)