Essay: I used AI to write fiction. Let’s talk about that.
I know this is a topic with lots of energy around it, so I want to be transparent.
Reading time: about 8 minutes
Some of you might not know exactly what I’m talking about, so to start: a brief introduction.
There is an AI model called GPT-3 that is very good at writing. One application of that model is in the software ChatGPT, where you can talk to the AI as if it were a bot and ask it to write something for you. It will usually spew forth a relatively well-written essay or article or poem or anything else. It’s the end of writing. Or at least, it’s the end of human writers. Or at least, it’s the end of jobs as writers.
The Internet is very excitable.
Don’t get me wrong. GPT-3 only works because someone built a training database for it, and that database is populated with the text from books, articles, ad copy, blogs, posts, and any other written thing that could be scraped off of millions of websites on the Internet. There’s a good bet many of those written things were copyrighted, and permission was not asked nor given. The ramifications of that are being litigated right now on image AIs that use the Stable Diffusion model. Stable Diffusion’s image AI can create images from simple text prompts. Like the GPT-3 writing model, the only way it can do that is with a training database of images. 2.3 billion images, to be exact, and once again, copyrighted images were used and permission was not asked or given. Litigation should happen, and it’s good that it’s happening sooner rather than later.
But back to the end of writing.
GPT-3’s magic is simply this: it has examined over 300 billion words that were arranged in ways that make sense to a human. It has figured out which word is statistically likely to come next. It has no clue what the words it is stringing together by probability mean. It does not know a preposition from an adjective. All it knows is how to recognize patterns well enough to math together sentences that look similar to ones it was trained on.
It’s the whizz-bang parlor-trick version of Autocomplete. I won’t say that it can’t make something new—it can, but only something that is slightly new. Because if it strays too far from the high probability of what comes next, what it spews out will stop making sense. Because it doesn’t know what “makes sense” even is. Having used an AI, I can promise you that even with 300 billion words studied, it will still produce complete nonsense half of the time. It’s not going to come up with new ideas. It’s not going to make interesting connections, or relate a human experience that wasn’t in its training data. It’s not going to write compelling characters driven by deep inner pain. All it can do is retread already existing sentences, and after the novelty wears off, I am betting that people are going to want to keep reading AI-created works about as much as they want to read ad copy.
AI-assisted writing—now, that is just fun.
Let’s talk about that difference for a moment. You can use AI to write for you. You give it an idea for a short article or essay (because there’s no AI out there yet that has an inborn need to express itself), and it will write out the idea from start to finish. And then you take that output and publish it as a finished piece. This is the vision of “the end of writing” that has writers very nervous. That readers could enter in a text prompt and an AI will vomit forth a bespoke novel just for them in under a minute. I can’t promise that day won’t ever come, but it is not today. GPT-3 can’t even analyze or output more than roughly 1500 words at a time, so there is no way for it to keep track of what happens in a novel (usually around 85,000 - 115,000 words), and no way to write more than about a scene at a time. This means that to put together a coherent narrative there needs to be lots of human intervention and prompting along the way.
AI-assisted writing means just that—the AI as a tool, not as creator. And as a tool, I’m loving it. Whenever I get stuck in a story, I ask the AI what it thinks should come next. In general, the answers it comes up with are painfully trite or completely overblown or make no sense whatsoever. I read them and think: oh, hell no! No way is that happening in my story! THIS is how it should go!!!
Oh look, no more stuck. Or the nonsense will end up as a bridge to an interesting idea, or a turn of phrase will give me a new insight. It’s like describing your story to someone else and they completely miss the point, which makes you want to explain it harder. The AI can be inspirationally bad. It can also be plain inspirational, when I have small details in the story like the style of a character’s shirt or the description of a lobby to write out.
The point to AI-assisted writing is to help smooth out the process. I never rely on the AI for the exact words I am going to use in my story. In most drafts I use less than 10% of the AI’s suggested words. I use the AI’s magic autocomplete fumbling like the terrible advice you get from a friend. You don’t take the advice, but upon reflection it was nice to talk to someone else about it, and that inspired new contemplation of the problem.
Why am I bringing all this up? Because AI is becoming more and more the topic of conversation in fiction writing circles. While I was looking for magazines to submit stories to, a number of them now state in their submission requirements that they will not accept stories written by AI or stories that were written with AI assistance. Of course, that is all on the honor code—there is often no way to tell. But it is one reason that helped me decide to keep all my writing on Substack for now—I don’t want to give up my AI writing buddy. Writing with the AI just feels—lighter. Less draining. It makes some of the small decisions for me, which gives me some energy back to get more done. And I think when it comes up with some absurd suggestions, it reminds me that my writing is not life or death. I can make the wrong decisions, pick ridiculous words, and have a draft that makes no sense, because it’s just a draft. It’s all fixable. See, look at what that crazy AI just wrote—mine’s better than that, at least.
I know that AI is not without flaws. Most of its flaws center around all Big Tech’s flaws: exploitation, a move fast and break everything philosophy, stealing other people’s work, and all the -isms. Anytime we as a society get on reigning in the tech giants, I’m there for it, I sign the petition, I vote, I donate.
I also can’t guarantee that writers won’t lose jobs to this AI, but I also don’t think that writing as a career is dead. I was a tech writer for 15 years, and throughout that entire time certain bosses and coworkers let me know clearly and regularly that they could do my job, just, you know, who would want to? All those people will prefer to pay an AI instead of pay a salary. Except getting the text correct and clear for anything that needs specificity is going to take a lot of skill and time prompting the AI and editing its output. So copywriters might spend a lot more time editing AI text than creating copy from scratch. And I could see some businesses just settling for whatever the AI puts out, and it will be interesting to see over time if that is their final choice. AI also can’t create or manage a knowledge base, or take notes, or conduct interviews, or give presentations, or create or follow style guides and templates. It can’t figure out what is important to convey and what isn’t.
If we can iron out some of the thornier issues, which we seem to be starting to do, AI could be a great tool for writers, just like spell check is. (Which we are still arguing over whether is good or bad for us, more than 30 years later, so it’s not like the all over the place reaction to this AI tool is a big surprise.) But I think it could do more good than harm. I have a dyslexic friend who is forced to craft emails as part of her job and it is the bane of her existence—AI can absolutely help with that.
So that’s why I use AI—for now. After the legal dust settles I’ll revisit the decision. Until then, I’m going to keep exploring what this new tool can help me do.
I also use Midjourney for (most) images on this Substack. This essay has gone on long enough though, so I’ll talk about my thoughts and feelings around image AIs soon.
"I use the AI’s magic autocomplete fumbling like the terrible advice you get from a friend. You don’t take the advice, but upon reflection it was nice to talk to someone else about it, and that inspired new contemplation of the problem."
I love this line and I absolutely will smurf it for use in conversations about AI-assisted code development
"The Internet is very excitable."
This is very funny and made me laugh.
Also, discourse on the Internet rarely leaves room for nuance. This is why someone can post the most innocuous thing about liking flowers or spending time with loved ones and it will manage to garner rage. I appreciate you sharing some nuance around AI writing programmes when almost everyone seems to be jumping on the "THING BAD" bandwagon.