Because fiction is a different genere and historically at least AI was much worse at it.
The big differences:
- fiction tends to be longer, AI struggles with making a satisfying coherent plot structure after a certain length.
- fiction tends to be subtler. You want characters to have nuance, shades of grey, symbolism, etc. Not everything should be shouted in your face. This is the opposite of writing persusaive literature where you are trying to convince someone of something.
There's money in copywriting, there's money in writing code, but long form creative writing? Almost an opposite of money.
Thus, very little incentive for AI labs to spend any effort on improving long form creative writing performance.
Most of tunes on this generation of AI systems are basic RLAF, and aimed at something like "punchy short form writing". The potential is largely unexplored.
I tried it once. Maybe 8-9 months back.I recall doing it soon after one travel, so that's why I can pinpoint it and this may be useful to figure which AI versions were available then.
The challenge is length and the context window. I had to own the plot since once it compacts, it has already lost the plot :)
I will give it pointers and tell where to go next. It will do it but either 1) stop after 3-4 paragraphs, or 2) write in a totally different direction than I expected. So I will nudge it to course correct, and it will do it. And then you keep iterating.
It was a good experiment for a couple of days. It saved me typing and it sometimes gave me lines of thinking that were interesting. But nothing that could be published.
But now that I think about it, I found a use case. If only GRRM could use its help :)
My point was, LLMs always generate fiction (bad fiction, but fiction nonetheless); what they generate is plausible, grammatical text that may have many points of congruence with the truth, but are not in any way constrained to do so. In short, fiction.
If they could reliably generate high quality non-fiction, that would be news.
You're not reading someone else's RP, though. Not with the applications that support this.
There are things like the character(s) you interact with, the initial setting, and some background information predefined, but the responses/evolution of the RP vary depending on things like the model used and the user's interactions.
It's somewhat like taking a D&D setting/scenario. Each group plays it differently.
Ah, the point I was trying to make was more akin to “why would I read your AI-generated story that you’ve posted as a static text/ebook trying to make a profit when I could do [essentially the scenario you described].”
To clarify, I meant that the draw of AI-written content would be exactly that interactivity - why read a story someone else has generated and posted in a static way when I could go and tailor one to my exact tastes instead. “No one wants to read AI generated texts” being appended with “that they haven’t gone and cooked up themselves”
Erotic content lowers the quality bar overall - or, maybe more proper to say, people are willing to put up with much worse quality, as long as it features content that arouses them. Settings, characters/pairings, sexualities, activities, dynamics, anything you can think of that someone might be into - media that’s a 5/10 normally becomes a 6 or 7/10 when it features your very specific underserved kink. And if you can generate it yourself, privately, practically for free?
Erotica is a whole separate class of smut than porn (in the cultural sense, that we understand the term “porn”). It’s literally like calling a book a movie.
It’s also, on the whole, a much more ethical and engaging form of smut. I don’t really have an issue with the choices of consenting adults, but the (video and image) pornography industry is so seldom that.
I've heard that word used to refer to (still) pictures, and in more casual/offhand use to refer to text. I don't understand it as being quite so restricted by media-type as you seem to.
- If you can generate something you're happy with, that's typically enough. External validation usually isn't a requirement for getting off to yourself, as opposed to other content where many of the primary benefits often are sharing
- Plenty of people are into very niche things. Sometimes almost nothing exists (or is discoverable) that exactly hits the spot. Generative AI lets you hit the spot that nothing else has
- AI is less complex than managing a real friendship. Even if people exist who can help you out with this, it can be difficult to arrange, or you lose access to them etc. AI doesn't lose interest in you, get busy with life stuff or decide you don't deserve a second chance.
Also sharing is still a huge benefit if you can manage it. Sometimes you will only find other people share your spot once you start posting the type of content they've been looking for. (Woe is me that FurAffinity unconditionally bans any LLM assistance now)
I had a sci-fi plot sketched out in my head, I had Codex give a whack at it. It was "ok" - not the worst content I've ever read, not the best. It was "sufficient."
my speculation is that not having written the original text means you don't have a mental history of all the decisions that went into writing it, and the context from being immersed in the work, making it a lot harder and less motivating to go write the next draft.
it was mostly just an experiment to see how well it could go. I didn't have a huge amount of motivation to keep going with it, I just wanted t osee if it could do it. Arguably you could flesh it out way more and have a legit work of fiction, but I only have so many hours in the day and was more interested in seeing if it could make something passable.
My take? It was passable. Over the last year I've read quite a bit less because I spend a lot of times looking at screens during the day and periodically get headaches, but I read something like 12-36 books a year depending on the year. Some of the fiction I read is good, others is... well, it's adequate.
From a purely "was it good enough to make money" standpoint, a book written exclusively with codex could probably sell copies. There are a lot of crappy authors out there too.
It's a very fun to go on an exploration of a theme! I had one where the technology was that people could go to sleep for free, but they had to pay to wake up, and when, and the economics and how society got rebuilt around that, and of course the hero trying to break the system.
I think I probably did 12 whole books before I realized no one else would read them, or that if they did, I would feel guilty. some of them are really good though even knowing they are purely plan-based expression!
This is interesting. We have seen the internet change many fields and democratize them. For instance, only a few media outlets produced news stories and analysis, the rest of us consumed it. Blogging changed that.
Only a few studios produced shows. With Youtube etc., many of the consumers could become producers themselves.
If I read this correctly, books and fiction are headed in the same direction.
I'm not sure about that comparison. For news and television your analogues of blogs and YouTube overcome a distribution bottleneck. Books have for a long time had a low barrier to entry for distribution and that fell further with the internet. There are mountains of amateur fiction and fanfiction online, requiring only an internet-connected device to produce and consume.
LLM-written fiction as explored in the study is generally not published at all. It's treated more like an externalized imagination, a loop of general ideas fed into and expanded on or filled in by the machine with statistical averages. It more closely resembles a sandbox game in my view, a type of media distinct from anything before it in form, and even more distinct in function in that media is generally understood to be a vector of communication between people, and this is instead highly individual.
Actually, it might be closer to say this is similar to a child playing pretend alone with their toys, except perhaps a bit less challenging in that creative roadblocks or narrative building is instantly abdicatable to the machine.
> Actually, it might be closer to say this is similar to a child playing pretend alone with their toys, except perhaps a bit less challenging in that creative roadblocks or narrative building is instantly abdicatable to the machine.
Wouldn't the obvious analogue be a video game? Especially one where you can edit the asset files (making your weapons super-strong, for instance?)
No, actually. In a video game the rules of engagement and the content you're engaging with are authored by other people. It's interactive media, but it's still a form of communication with boundaries and intent behind it. You can break those bounds with cheats or tools in the same way you can mark up a novel and write in your own ending with whiteout and a pen, but you're still fundamentally engaging with someone's creation.
An LLM-authored interactive fiction is a sandbox with no author, with no boundaries and no intent. It's a canvas, and all limitations on content are self-imposed. It's closer to making a game than to playing one, but there is no game in the end that you can share. You're left with a transcript that only really appeals to you. It is not the game. The game fundamentally lives in the player's imagination, and the LLM is there to ease the realization of that imagination and, if I'm being uncharitable, do the hard part of synthesizing ideas and impulses into something semi-coherent, limiting the potential for the player to grow and develop creative skills.
Web fiction, freely produced and distributed by "consumers" already existed prior to 2019. There are thousands of novels you can read that people produced for their own enjoyment, some even managed to make money off of it via patreon or Amazon's Kindle direct publishing.
This is slightly different than web fiction. Text generation is arguably the cheapest and most “ready” medium for content production in the current AI wave.
You can speak a world into existence, entirely customized to you’re preferences, and interact with it.
> I don't really see how AI 'democratizes' fiction and book writing?
When people say "AI will democratise $FOO" they don't mean it the way it is usually understood: people with skill and motivation can now produce $FOO.
What they mean is that people with no skill and no motivation can now pretend that they produced $FOO.
Remember the "I'm the ideas guy" people who would ask you to build a complete system and offer 1% equity because "the ideas are the hard part"? Well, the future is going to be filled with "ideas guys", making it even harder to find signal in the noise.
You need to learn to be able to say no to people as a life skill, first off, but if the ideas guy has vibecoded a mountain of code that they're trying to ship, but has hit the vibe wall, and has money to pay you to fix it, say yes.
So fans have fan fiction because they love certain worlds and characters and want more of it, or to riff on the base to take stories in new directions. We have things like TV tropes so it is no small stretch to, say, generate sitcom-type stories from Harry Potter or GoT that can continue forever. But where does all of this lead?
Music has sinilar issues, but I think of this like the compression/loudness issue in music production. Everything gets amplified so the range of everything is compressed. And then it gets boring and people slowly jump ship for something else.
I think there will be a wave of AI slop that improves and might actually be exciting on some vector, but we will get bored eventually. Humans crave newness, even if that new thing is worse in exactly the ways that defined good previously, like punk in response to complex rock albums.
AI can combine ideas in interesting ways, but it is by design a predictor of what is most likely. This is directly odds with the concept of newness (and arguably, human-ness) which is baked in to what we consider interesting and relevant.
Humans may crave newness, but it only needs to be new to them and also still mostly recognizably familiar. So artists who spent too much time steeped in the output of prior artists ended up with Dadaist literature and abstract paintings, while the average person still enjoys conventional novels and paintings that look like something.
I'm not sure whether Harry Potter pioneered the "school + wizards" mashup, but the school has many aspects you would expect of a school (save for the wizards) and the wizards have many aspects familiar from previous fictional wizards (save for the school part.) The novelty comes from combining the concepts in more or less straightforward ways.
Even a writing AI that is only capable of taking an existing plot and adapting it for a different setting could produce enough reading material to last a lifetime by taking all existing novels in the public domain and putting each in a different random mashup for modern audiences. And I think many people wouldn't tire of this at all.
Yes, they are. People have to find ways to make money to pay rent, bills and eat, particularly when their resumes are not as densely packed as others. This is one way to do it. Wish I could do it, too :)
LLMs suck at creative writing, all their edges get smoothed the fuck out of with RLHF for (brand) safety, so they write like they have a HR lady in the room with them, scrutinizing every sentence for adherence to the set of values very few people honestly hold. if you feed Claude/GPT/Gemini 100 chapters of ASOIAF and prompt it to continue, the story and the prose will degenerate into sterile purple slop within a few pages.
Interesting analysis, but data "was collected between April 2023 and May 2024", so this predates even the release of GPT 4o.
SillyTavern and the size of like Character.ai at its peak shows pretty obviously that people are doing exactly this!
> Are readers generating fiction with AI models?
Why not? Journalists, lawyers and pundits of various stripes are already doing it. Why shouldn't readers?
Because fiction is a different genere and historically at least AI was much worse at it.
The big differences:
- fiction tends to be longer, AI struggles with making a satisfying coherent plot structure after a certain length.
- fiction tends to be subtler. You want characters to have nuance, shades of grey, symbolism, etc. Not everything should be shouted in your face. This is the opposite of writing persusaive literature where you are trying to convince someone of something.
And also: there's no money in writing fiction.
There's money in copywriting, there's money in writing code, but long form creative writing? Almost an opposite of money.
Thus, very little incentive for AI labs to spend any effort on improving long form creative writing performance.
Most of tunes on this generation of AI systems are basic RLAF, and aimed at something like "punchy short form writing". The potential is largely unexplored.
I tried it once. Maybe 8-9 months back.I recall doing it soon after one travel, so that's why I can pinpoint it and this may be useful to figure which AI versions were available then.
The challenge is length and the context window. I had to own the plot since once it compacts, it has already lost the plot :)
I will give it pointers and tell where to go next. It will do it but either 1) stop after 3-4 paragraphs, or 2) write in a totally different direction than I expected. So I will nudge it to course correct, and it will do it. And then you keep iterating.
It was a good experiment for a couple of days. It saved me typing and it sometimes gave me lines of thinking that were interesting. But nothing that could be published.
But now that I think about it, I found a use case. If only GRRM could use its help :)
My point was, LLMs always generate fiction (bad fiction, but fiction nonetheless); what they generate is plausible, grammatical text that may have many points of congruence with the truth, but are not in any way constrained to do so. In short, fiction.
If they could reliably generate high quality non-fiction, that would be news.
"wrong" or "random" is not the same as "fiction"
"non-fiction" doesn't mean "correct".
https://www.merriam-webster.com/simple/fiction
https://www.merriam-webster.com/dictionary/joke
Words can have multiple senses, and in context i agree with the person claiming you are using the word wrong, given the context.
I suspect that those other groups are largely not trying to generate fiction.
I didn't see the word "try" in the headline, did you?
Because nobody wants to read AI generated text?
No, nobody wants to read someone else's AI-generated text. Some people don't mind reading text generated to their own specifications.
This is quite an important point.
Tell that to the LLM RP crowd.
I think you lose that draw of interactivity when you’re essentially reading someone else’s RP, though.
You're not reading someone else's RP, though. Not with the applications that support this.
There are things like the character(s) you interact with, the initial setting, and some background information predefined, but the responses/evolution of the RP vary depending on things like the model used and the user's interactions.
It's somewhat like taking a D&D setting/scenario. Each group plays it differently.
Ah, the point I was trying to make was more akin to “why would I read your AI-generated story that you’ve posted as a static text/ebook trying to make a profit when I could do [essentially the scenario you described].”
As someone with extensive experience with this: no, you don't.
As much as it's nice to know another shares my interest (and will partake in it with me), I can't always get that.
When I can't get that, LLMs are not a worthless substitute, especially if what I actually want is just to play out a situation.
I want to act and then get back a description of reaction. LLMs can describe reaction, therefore I find value in them.
Like how in D&D you go to do whatever and the dungeon master tells you what happens in response. LLMs fit there for me.
They don't fill the void of social interaction or anything like that, but I don't strictly need that here.
Take a look at services like AI Dungeon to see the kind of thing I'm talking about.
It doesn't really replace social anything, but it's kind of like text adventure. It's plenty interactive enough.
To clarify, I meant that the draw of AI-written content would be exactly that interactivity - why read a story someone else has generated and posted in a static way when I could go and tailor one to my exact tastes instead. “No one wants to read AI generated texts” being appended with “that they haven’t gone and cooked up themselves”
I think it's pretty clear from the discussed study that _someone_ wants to read that.
You do not represent everyone
Yeah, but i do, and i agree with the parent comment
[dead]
> We show that users especially gravitate toward fanfiction and erotica
Why am i not surprised its primarily porn
Erotic content lowers the quality bar overall - or, maybe more proper to say, people are willing to put up with much worse quality, as long as it features content that arouses them. Settings, characters/pairings, sexualities, activities, dynamics, anything you can think of that someone might be into - media that’s a 5/10 normally becomes a 6 or 7/10 when it features your very specific underserved kink. And if you can generate it yourself, privately, practically for free?
That’s why you’re not surprised.
Erotica is a whole separate class of smut than porn (in the cultural sense, that we understand the term “porn”). It’s literally like calling a book a movie.
It’s also, on the whole, a much more ethical and engaging form of smut. I don’t really have an issue with the choices of consenting adults, but the (video and image) pornography industry is so seldom that.
I've heard that word used to refer to (still) pictures, and in more casual/offhand use to refer to text. I don't understand it as being quite so restricted by media-type as you seem to.
- If you can generate something you're happy with, that's typically enough. External validation usually isn't a requirement for getting off to yourself, as opposed to other content where many of the primary benefits often are sharing
- Plenty of people are into very niche things. Sometimes almost nothing exists (or is discoverable) that exactly hits the spot. Generative AI lets you hit the spot that nothing else has
- AI is less complex than managing a real friendship. Even if people exist who can help you out with this, it can be difficult to arrange, or you lose access to them etc. AI doesn't lose interest in you, get busy with life stuff or decide you don't deserve a second chance.
Also sharing is still a huge benefit if you can manage it. Sometimes you will only find other people share your spot once you start posting the type of content they've been looking for. (Woe is me that FurAffinity unconditionally bans any LLM assistance now)
Fwiw, the unconditionally bans can almost be viewed as 'proof of human' when so many bots are running on autopilot on places like DA. UwU
I don't mind bans on fully generated content, I just feel it's a bit unfair to ban any usage of generative AI anywhere.
I had a sci-fi plot sketched out in my head, I had Codex give a whack at it. It was "ok" - not the worst content I've ever read, not the best. It was "sufficient."
But now you have a starting point to edit from. Just find a bit more motivation to continue writing...
my speculation is that not having written the original text means you don't have a mental history of all the decisions that went into writing it, and the context from being immersed in the work, making it a lot harder and less motivating to go write the next draft.
it was mostly just an experiment to see how well it could go. I didn't have a huge amount of motivation to keep going with it, I just wanted t osee if it could do it. Arguably you could flesh it out way more and have a legit work of fiction, but I only have so many hours in the day and was more interested in seeing if it could make something passable.
My take? It was passable. Over the last year I've read quite a bit less because I spend a lot of times looking at screens during the day and periodically get headaches, but I read something like 12-36 books a year depending on the year. Some of the fiction I read is good, others is... well, it's adequate.
From a purely "was it good enough to make money" standpoint, a book written exclusively with codex could probably sell copies. There are a lot of crappy authors out there too.
Probably would be good to write up a TTRPG adventure, though!
It's a very fun to go on an exploration of a theme! I had one where the technology was that people could go to sleep for free, but they had to pay to wake up, and when, and the economics and how society got rebuilt around that, and of course the hero trying to break the system.
Don’t outsource your brain. Give your brain a whack at it.
[dead]
I think I probably did 12 whole books before I realized no one else would read them, or that if they did, I would feel guilty. some of them are really good though even knowing they are purely plan-based expression!
Language game solipsism
This is interesting. We have seen the internet change many fields and democratize them. For instance, only a few media outlets produced news stories and analysis, the rest of us consumed it. Blogging changed that.
Only a few studios produced shows. With Youtube etc., many of the consumers could become producers themselves.
If I read this correctly, books and fiction are headed in the same direction.
I'm not sure about that comparison. For news and television your analogues of blogs and YouTube overcome a distribution bottleneck. Books have for a long time had a low barrier to entry for distribution and that fell further with the internet. There are mountains of amateur fiction and fanfiction online, requiring only an internet-connected device to produce and consume.
LLM-written fiction as explored in the study is generally not published at all. It's treated more like an externalized imagination, a loop of general ideas fed into and expanded on or filled in by the machine with statistical averages. It more closely resembles a sandbox game in my view, a type of media distinct from anything before it in form, and even more distinct in function in that media is generally understood to be a vector of communication between people, and this is instead highly individual.
Actually, it might be closer to say this is similar to a child playing pretend alone with their toys, except perhaps a bit less challenging in that creative roadblocks or narrative building is instantly abdicatable to the machine.
> Actually, it might be closer to say this is similar to a child playing pretend alone with their toys, except perhaps a bit less challenging in that creative roadblocks or narrative building is instantly abdicatable to the machine.
Wouldn't the obvious analogue be a video game? Especially one where you can edit the asset files (making your weapons super-strong, for instance?)
No, actually. In a video game the rules of engagement and the content you're engaging with are authored by other people. It's interactive media, but it's still a form of communication with boundaries and intent behind it. You can break those bounds with cheats or tools in the same way you can mark up a novel and write in your own ending with whiteout and a pen, but you're still fundamentally engaging with someone's creation.
An LLM-authored interactive fiction is a sandbox with no author, with no boundaries and no intent. It's a canvas, and all limitations on content are self-imposed. It's closer to making a game than to playing one, but there is no game in the end that you can share. You're left with a transcript that only really appeals to you. It is not the game. The game fundamentally lives in the player's imagination, and the LLM is there to ease the realization of that imagination and, if I'm being uncharitable, do the hard part of synthesizing ideas and impulses into something semi-coherent, limiting the potential for the player to grow and develop creative skills.
Web fiction, freely produced and distributed by "consumers" already existed prior to 2019. There are thousands of novels you can read that people produced for their own enjoyment, some even managed to make money off of it via patreon or Amazon's Kindle direct publishing.
This is slightly different than web fiction. Text generation is arguably the cheapest and most “ready” medium for content production in the current AI wave.
You can speak a world into existence, entirely customized to you’re preferences, and interact with it.
You should suggest some!
I recommend Super Supportive[0].
[0]: https://www.royalroad.com/fiction/63759/super-supportive
Why would you need to "democratize" fiction? There's literally zero cost of entry already. Fiction probably needs more gate keeping.
> If I read this correctly, books and fiction are headed in the same direction.
I don't really see how AI 'democratizes' fiction and book writing?
> I don't really see how AI 'democratizes' fiction and book writing?
When people say "AI will democratise $FOO" they don't mean it the way it is usually understood: people with skill and motivation can now produce $FOO.
What they mean is that people with no skill and no motivation can now pretend that they produced $FOO.
Remember the "I'm the ideas guy" people who would ask you to build a complete system and offer 1% equity because "the ideas are the hard part"? Well, the future is going to be filled with "ideas guys", making it even harder to find signal in the noise.
You need to learn to be able to say no to people as a life skill, first off, but if the ideas guy has vibecoded a mountain of code that they're trying to ship, but has hit the vibe wall, and has money to pay you to fix it, say yes.
Books and fiction were some of the first to be democratized as a result of the internet.
So fans have fan fiction because they love certain worlds and characters and want more of it, or to riff on the base to take stories in new directions. We have things like TV tropes so it is no small stretch to, say, generate sitcom-type stories from Harry Potter or GoT that can continue forever. But where does all of this lead?
Music has sinilar issues, but I think of this like the compression/loudness issue in music production. Everything gets amplified so the range of everything is compressed. And then it gets boring and people slowly jump ship for something else.
I think there will be a wave of AI slop that improves and might actually be exciting on some vector, but we will get bored eventually. Humans crave newness, even if that new thing is worse in exactly the ways that defined good previously, like punk in response to complex rock albums.
AI can combine ideas in interesting ways, but it is by design a predictor of what is most likely. This is directly odds with the concept of newness (and arguably, human-ness) which is baked in to what we consider interesting and relevant.
Humans may crave newness, but it only needs to be new to them and also still mostly recognizably familiar. So artists who spent too much time steeped in the output of prior artists ended up with Dadaist literature and abstract paintings, while the average person still enjoys conventional novels and paintings that look like something.
I'm not sure whether Harry Potter pioneered the "school + wizards" mashup, but the school has many aspects you would expect of a school (save for the wizards) and the wizards have many aspects familiar from previous fictional wizards (save for the school part.) The novelty comes from combining the concepts in more or less straightforward ways.
Even a writing AI that is only capable of taking an existing plot and adapting it for a different setting could produce enough reading material to last a lifetime by taking all existing novels in the public domain and putting each in a different random mashup for modern audiences. And I think many people wouldn't tire of this at all.
I don't get your point. Do people not get bored and move on to something else one way or the other? What exactly is different?
For fun I wrote three chapters, and let opus write the 4th. It was god fucking awful. No amount of prompting could fix it.
Yes, they are. People have to find ways to make money to pay rent, bills and eat, particularly when their resumes are not as densely packed as others. This is one way to do it. Wish I could do it, too :)
LLMs suck at creative writing, all their edges get smoothed the fuck out of with RLHF for (brand) safety, so they write like they have a HR lady in the room with them, scrutinizing every sentence for adherence to the set of values very few people honestly hold. if you feed Claude/GPT/Gemini 100 chapters of ASOIAF and prompt it to continue, the story and the prose will degenerate into sterile purple slop within a few pages.
[flagged]
[flagged]