Warning: some of this is going to sound a lot like snobbery, and maybe it is. A couple of people have told me it isn’t, that it’s insight built up from years of writing and reviewing. Judge for yourself.

About 10 years ago, I read about an elderly woman on what was then Twitter who spent her days posting and reposting huge volumes of material in support of her favoured political candidate (doesn’t matter who that was, not the point). Twitter’s automatic moderation kept banning her as a bot.

I remember thinking at the time, “Well, isn’t she a bot? A bot implemented in the form of an elderly lady, sure, rather than as software, but she’s acting exactly like a bot. The auto-moderation is right to treat her as one.”

I’m reminded of that because I’ve been thinking about the coming impact of AI on the creative arts.

I recently read a piece on The Conversation about how, if you feed a large language model (LLM) a prompt and then get it to alternate between turning text into images and images into text for a few iterations, you will quickly get bland, generic, meaningless images without any relationship to the original prompt.

This isn’t surprising. After all, LLMs are not actually intelligent, artificially or otherwise. They’re basically sophisticated autocorrect. They’re trained to predict what most probably comes next, or the most likely response to a prompt. Of course they’re going to produce something bland and expected. That’s what they’re built to do.

(My friend who knows a great deal more about AI than I do says this is so much of an oversimplification that it almost isn’t true, and also the study is flawed if you think about it even a little bit. I’ll write a follow-up when I get time, discussing this. But the main point here is that LLMs have at least a tendency to return expected content, and anything stronger than that statement is not essential to the following discussion.)

Which is potentially bad news (or… is it?) for the people who have, for a few years now, been producing bland, expected fiction “written to market” – in other words, written to be similar to already existing properties that are successful. And also for Hollywood, which is producing a parade of increasingly bland, formulaic remakes and sequels. I don’t listen to current music, but I’m informed it’s going through the same blandification process, thus perfectly setting up for LLMs to come in and take over. (I don’t think this is the result of a conspiracy, to be clear.)

And the thing is, LLMs are often very good at doing this kind of thing. And they can punctuate a lot better than most of the people who are churning out fiction rapidly from box mix and selling it to people who want the same thing over and over again. I sometimes refer to this as “extruded fiction product” or, if it’s done less completely formulaically, “If you like this sort of thing, this is definitely one.” Just as the elderly lady was acting like a bot, these people have been acting like LLMs, meaning that it’s potentially easy for LLMs to do what they do.

Though, having said that, it’s also emerging from research into the use of LLMs in software development that the people who are getting the most gains are the people who have the best engineering discipline in the first place, who are using the LLM for the tedious parts and carefully checking its output. So perhaps we can’t dodge that work after all – or, if we do, we need to be aware that the outcome won’t be great. But the people I’m mostly talking about are the people who already didn’t care much about quality, who put a higher value on velocity. They will now be able to produce more low-quality fiction (or code, or visual art, or video) than ever before. I haven’t been on Facebook for years, happily, but I’m told it’s already overrun with low-quality, low-effort AI images, which the platform doesn’t do much to discourage because they get “engagement”. I have to believe that if people buy crappy LLM-produced fiction on Amazon, Amazon isn’t going to try very hard to stop them.

On the flip side, a recent study reports that, while on a standard test of “linguistic creativity” LLMs do better than the average human, they do significantly worse than the average of the top 50% of humans, and the gap widens if you look at the top 25% or top 10%. Even if you turn up the randomness in the model to its maximum, so that it’s producing less “expected” outputs, more than 25% of humans still score above the LLM. Yes, the models will continue to improve, but I believe there’s an upper bound just by the nature of what an LLM is – and you rapidly hit a law of diminishing returns as you push towards larger models that take more factors into account or can retain a longer window of context, such as you would need in a novel. LLMs can’t write whole novel-length books yet, not coherently at least, and perhaps will never be able to, barring some advancement in computing that’s beyond our current prediction horizon [edit: this is also incorrect]. (What they can do is produce 95% of some of the copyrighted work they’ve been trained on, verbatim, which, in at least the case of Meta’s AI, incudes some of my books, via a pirate site. More on that in a future post, but that would certainly make it easier to produce new versions of those same books with a few names changed.)

Also, increased randomness doesn’t necessarily produce what humans would rate as better outputs, just more unusual ones. Top human creatives not only have the skill to generate unexpected ideas, they also have the skill to distinguish a ridiculous idea from one with potential to be great. Not to mention that they have a depth of insight into humanity that is part of what lifts their creative works above the ordinary.

Over the past couple of years, I’ve been reading a lot of old fiction from Project Gutenberg, in part because it’s hard for me to find enough fiction being produced today that isn’t bland and generic, and I’ve discovered something that will probably not surprise you even a little bit. A lot of the fiction being produced in, say, the 1920s was pretty much alike, shallow, conventional, and not very well written. The enduring classics, the ones that are still widely read today, that are reprinted (often with minimal editing from a bad scan) by major publishers, are the exceptions, the ones that were doing something different, something original, something inimitable.

So, while I have all kinds of concerns about AI (more on this in an upcoming article), I do have a form of hope, and it’s this. I hope that we’ll start to see a divide develop. On the one hand will be the people who are now using AI to do what they used to do themselves – produce bland, cloney books that are almost indistinguishable from each other to feed a particular appetite in the market for something recognisable and unchallenging. I suspect the quality may even rise a bit, though the coherence may drop. And on the other hand will be people who are exercising human creativity that can’t be imitated by an algorithm that seeks to predict what’s most likely to come next.

The key thing will be how easy it will be to distinguish the two in the market. If there’s a “not made with AI” label of some kind that people can use (which itself will be fraught with controversy as people inevitably try to game it, not to mention how you define “no AI” when it’s built into so many tools), the divide will be more obvious, and people will be more easily able to buy the kind of fiction they prefer. Most will prefer the expected kind, because that’s always been the case since cheap fiction became technologically possible. But some will prefer the unexpected kind, and they will potentially have a discovery problem, of which more in my promised follow-up.

I suspect they’re the ones who are already ignoring the dozens of similar books about, for example, a young (or middle-aged) woman who returns to the small town where she grew up to deal with an inheritance, discovers she has a connection to the supernatural, gets improbably accused of a crime, and works to solve it along with her cute, chatty familiar, her completely subservient best friend and the hot guy who’s a police officer and/or werewolf.

It is possible to write even that story in a fresh way – I’ve seen it done (by Tim Pratt in Heirs of Grace – link is to my review). Not many people can do it, though. An LLM can’t do it and, by the nature of LLMs, probably won’t ever be able to. It can only write something close to the expected version, because that’s what an LLM is.

Now, there is another problem inherent here, and it’s the problem of development. In order to get good at any skill, you have to go through a period where you’re not good at it, and I suspect that if an LLM is better than the average person at many skills (programming, visual art, writing), a lot of people will never bother to go through that period of learning for those skills. They won’t see the point in producing work that, to begin with, will be worse than what the LLM could do.

Which means they will never get better than the LLM. And we need people who are better at those things than an LLM can ever be.

Still and all, I’m not very good at making physical things. In fact, I’m notably bad at some aspects of it. That doesn’t stop me from trying, when there’s a specific thing I can’t just order, or when someone is going to charge me ten times what the materials cost to do, yes, a better job, but not a ten-times-better job, or when I just want to explore an idea that nobody else seems to have had. And in an age of manufacturing now several centuries old, there are still plenty of people who enjoy the process of making things by hand for its own sake, and people who will buy those things at a premium because they’re made by somebody rather than something.

People who are going to be the truly great programmers, artists or writers are going to persevere through the period of not being good at it because they have a vision of being better than average, and maybe even because they enjoy the process itself, and they may even start out at a higher level than the starting point for most people. Meanwhile, the people who were never going to be very good at those skills have access to a tool that can help them be better, just like my power saw helps me make more accurate cuts more easily than I could make them by hand.

There will potentially be losses in the middle – the people who don’t persevere to get as good as they could be, and the people who were always going to be average or no more than a little bit above it, whatever their ambitions. They will eventually be forced out of the market by people who have no ambition to be more than average and are willing to crank a handle in order to make a few dollars, because they never had any particular love of either the process or the product for its own sake. I’m certainly not saying that the impact of LLMs on creativity is going to be an unmixed benefit, and maybe it won’t be much of a benefit at all. Perhaps it will be an overall detriment.

But it isn’t going to bring human creativity to a halt, any more than the much-lamented “tsunami of crap” of the early self-publishing era meant that there were no good self-published books. After all, 90% of everything is crap; Sturgeon’s Law is universal, at least wherever humans are involved (even if only as toolbuilders). But there is still that other 10%, and it’s not going away.

(Brandon Sanderson has a great video which makes some of the same points as I make above, some that are in contrast and some that go beyond what I’ve said. In particular, he expands on the idea of the worth of the process, not only for the art but also, crucially, for the artist. He’s also one of the examples I point to when I talk about someone who’s creating something exceptional in part by not doing what’s expected. And as he explains in the video, he got to that point by being willing to write something that sucked, and learning from that so that the next thing sucked less, and eventually reaching the point where his sixth thing didn’t suck much at all and was publishable.)