(Note: in this piece, I link to a lot of websites that I’ve played with at different times. This is to allow the reader to make up their own minds about these things. No endorsement is intended, except of course when I am linking to my own books. If I didn’t like them, I wouldn’t publish them.)

First off, what we see today is not truly artificial intelligence, in the sense of “artificial sapient beings capable of exercising judgment and choice.” ChatGPT, Stable Diffusion and their many cousins and descendants are Large Learning Models (LLMs), software that has been programmed to extrapolate statistical information from the dataset it is given and offer randomized responses to human commands based on the dataset and the extrapolations. The results are only as good as the initial programming, and the dataset. And in terms of output, they’re only vastly more complicated versions of the random generators available at Seventh Sanctum and similar websites for more than a decade. Here are the situations in which I personally have used LLMs:

–Midjourney: I make pictures with this for the fun of it, mostly. I read somewhere that basically the monthly plan is the equivalent of three streaming TV services, and I feel like I get that level of entertainment out of it. (Not really a tribute to the quality of content on the streaming services, is it?) I made my last cover (Wolf’s Trail, currently available in Kindle Unlimited) with Midjourney and a font from Creative Fabrica and spent some hours tweaking image and text and putting it all together in paint.net. I used Midjourney because I wasn’t feeling the premades out there, was not going to be able to afford a $500+ custom design, and I already had Midjourney and knew more or less how to use it. I understand living artists’ concerns about their artwork being used for training purposes without their consent, and I did not use artist names to prompt Midjourney in creating the art used for Wolf’s Trail.

-transcription of dictated work by software “branded” as AI: I use dictation a fair amount, so I’ve played with a lot of different options for this and it would require another post to cover my thoughts. The short version is that they aren’t functionally different from the existing “Speech-To-Text” options from Microsoft, Google, etc, but some seem to get better results than others. All speech-to-text requires significant human correction; to the point where “cleaned up” transcription is in my opinion a different draft from raw transcription. This is where the “only a tool” status of these programs is most clear: a human is offering extensive initial input, the software is supposed to recreate the input in a new format as accurately as possible, which it tends to fail at, and then the human takes over again and finishes the job.

–Sudowrite: I played briefly with the free sample of this for brainstorming purposes during a moment last year when I was kind of stuck on Wolf’s Trail. Its actual prose was not useful for my purposes, I just went away thinking: “Well, yeah, it’s right about me needing to describe the town of Altstadt a little more and the countermeasures the locals are taking against the monsters, before Maxim leaves town to go to the long barrow.” So, I wrote that part myself, but the LLM helped get me unblocked mentally. There are people who use Sudowrite to write whole novels, but my main takeaway from the tutorials is that it looks like more work than doing it yourself.

–Kindlepreneur Book Description Generator: I experimented with this when I was looking to tweak my blurbs for the Jaiya Series and the Ancestors of Jaiya Series. Didn’t end up implementing any changes, or only minor wording tweaks if I remember correctly.

–Bard: The Vella version of Wolf’s Trail had a very basic blurb: two sentences from Chloe’s POV, two from Maxim’s, all 3rd person even though the novel itself is in first. I fed Bard the Chloe sentences from the Vella blurb and told it to rewrite them in first person. It spewed a ton of not very useful stuff at me. I added “and keep it short.” Found a phrase I could use with some rewording. Repeated the regular and “keep it short” prompts for the Maxim sentences. Got another phrase I could use with some rewording. Built a new blurb that incorporated those two phrases. End result: 172 words, 16 of them from Bard (around 9% of total blurb). Bard generated 2543 words in total; of which 0.6% were useful to me.

In conclusion, I really don’t see how using a LLM for writing a full novel would work. Basically it would be taking the (somewhat) fun part of writing away – the whole process of meeting the characters and enjoying their adventures along with them – and just leave me with editing/revision (much less fun). Most of my novels are in the 50000 word range. If 0.6% of Bard’s word output is usable for my purposes (based on the blurb experiment above), then I would need to have it generate something like 83000 words (probably prompting it on a scene by scene basis) that I would have to whittle down to a coherent 50000 word novel. Really not fun, especially since the entire reason I’m writing in the first place is because I’m dissatisfied with a lot of the stuff out there…the same stuff that probably forms part of the dataset for these LLMs. At a competition level, I’m already spectacularly unsuccessful compared to anyone who can write to market and release rapidly (or achieve the same results using human ghostwriters). I don’t really see how competing against author brands whose ghostwriters are LLMs would make my situation as an author worse. It’s just more of what’s been happening all along.

LLMs for image-making…that’s another matter. Visual art allows more room for surreal and non-rational content, which means that image-oriented LLMs are at less of a disadvantage relative to human artists. I feel like human artists are going to need to be much more proactive in protecting their creations going forward. People like me, who need commercial art done at rare intervals and can’t justify the cost and potential personal drama(1) of commissioning custom work, should probably avoid using Midjourney to imitate artists whose copyrights are still in effect. At least until the courts have settled the question.

(1) I personally haven’t had bad experiences, but plenty of authors have complained about artists who ghost them or fail to deliver on time.