On AI slop...

They aren't breaking any rules though, so I'm not sure what really can be done., and if the rules were changed where do you draw the line? One thing I would like to see is the algorithm for trending stories to be altered so it doesn't put quite so much of an emphasis on how many chapters are being added to a story regardless of whether they are getting much engagement or not. Somoene who writes fifty chapters which get one like each seems to be considered to have much more heat than someone who writes one chapter that gets fifty likes whereas I would argue it should probably be the other way around. That also very obviously favours AI authors who can often produce crazy amounts of chapters in a short space of time which human writers aren't going to be able to keep up with. The other thing the algorithm seems to strongly favour, as far as I can tell, is new stories but I quite approve of that since it is good to give those a boost while they are getting off the ground.
 
Last edited:
I for one would prefer there to be *no* AI slop at all, but that is clearly never going to happen because the RP bots are advertised here so it would be silly to not let the slop also be part of the site.
 
Marking AI work (as on DeviantArt) would also be my preferred solution. My one proviso would be that it would need to be clear we are talking about the text only. I'm fine with there being a separate box to tick for AI images but those of us who write our own text but use AI images should be able to be clear that the writing (the most important part on a literature site) is still our own even if we use AI tools for illustrative purposes.
 
I spent some time checking chapters of suspicious accounts on online AI-checking tools. Tools say those are fine, I highly doubt it...
So, how do you mark something as AI? Self-reporting? Opinion of a human moderator?

In any case, the problem is spam. Not the method of producing spam. One of the methods of fighting spam is ignoring it
 
One of the methods of fighting spam is ignoring it
This doesn't work, because the only way to detect bad writing is to read it.

That's true of the old kind of human-crafted spam, too, of course, but that was (and is) trivially easy to detect. Think all lower case, no punctuation, filled with changes in tense and point of view, littered with speling errors, paragraphs of character introductions before the plot even begins. You can dismiss stories like that after reading just 1 or 2 sentences. It's low-investment. 5 seconds, then move on to the next. There are other tells too, like chapter length, number of chapters, lack of cover image, and so on. In the end, they're mildly annoying, but no more than that.

LLM-generated text, on the other hand, is not bad in such an easily-detectable way. You (or at least I!) have to actually read multiple paragraphs to figure it out. It's that very time investment that pisses me off when I realize what I'm reading. And unlike with bad human writing, the tells change all the time, so it's a continuing arms race between LLMs and people who do not want to read LLM-generated text.

And since there's so much of it, looking for a new story to write becomes a much larger hassle than it was 5 years ago. Not just 20 times 5 seconds of minor annoyances plus 3 hours of engaged one-handed reading, but now also 10 times 5 minutes before I run into the third "not just X—but Y" in four paragraphs and the alarm bells finally start ringing.

In any case, the problem is spam. Not the method of producing spam.
Quantity has a quality all of its own, but it works the other way around as well: a mountain of hay obscures even the most beautifully crafted needle.

-----

I agree it's a hard problem, though (obviously, given the above :P). Voluntary marking would be a useful (and easy!) first step.
 
I don't think AI can write better than a good human author (yet). But it is great for finding typos/grammar issues and pointing out weaknesses in your writing. Probably we should be using AI more for that.
 
This doesn't work, because the only way to detect bad writing is to read it.
If I see a user who posted ~100 stories, ~300 chapters, ~1M words in one year since registration, then I can conclude that is not a genuine author. And while there is a possibility that the author is just obsessed or has a basement full of slaves typing those, AI slop machine is far more likely. And this kind of AI use annoys me a lot

The thing is, this user doesn't seem to break any of the existing rules.
 
If I see a user who posted ~100 stories, ~300 chapters, ~1M words in one year since registration, then I can conclude that is not a genuine author. And while there is a possibility that the author is just obsessed or has a basement full of slaves typing those, AI slop machine is far more likely. And this kind of AI use annoys me a lot

The thing is, this user doesn't seem to break any of the existing rules.
I now know who you are referring to. But honestly it is really easy to just not read their stuff. The bigger concern for me as others have expressed is when you read several chapters into something before the penny drops that it's AI, in my experience usually through some major continuity error that a human author would have been unlkely to have made. That's the kind of thing that also implies someone isn't writing with AI they are literally just generating chapters and posting them quite posibly unread.
 
If I see a user who posted ~100 stories, ~300 chapters, ~1M words in one year since registration, then I can conclude that is not a genuine author. And while there is a possibility that the author is just obsessed or has a basement full of slaves typing those, AI slop machine is far more likely. And this kind of AI use annoys me a lot

The thing is, this user doesn't seem to break any of the existing rules.
I agree. Thankfully, extreme outliers is like this are usually pretty obvious. Pre-AI there were a few human authors that posted a ton of really bad, tiny stories, and I ended up writing a small userscript to hide their stories. If it were just one AI generator, or a handful, I could add them to the hide list and move on with my life.

But the LLM problem is much wider than that, and most AI generators aren't so easy to detect. In a more general sense, if a story looks interesting at first blush, then to find out if the author is posting a suspicious volume of ConTEnT, I first have to click on the username, then collate several pieces of data across multiple tabs and do advanced math (division) to see if their output per day or month or year is implausibly high. And of course, most AI generators don't dump a hundred thousand words per month onto the site. There's significant overlap between a few days of AI prompting a month and NaNoWriMo-level human productivity, meaning I still have to do the other thing too.
 
I documented with full disclosure my own brief experimentation with AI in "Wishes" (a spur of the freeuse branch). I ran it for a bit, posted a few lightly edited things for people to see (with commentary), then rewrote most of it using my original outlines I had fed the AI (and the occasional probably-accidental good thing the AI happened to spit out) and went back to writing manually. My basic conclusion was that, at least at the locally runnable level, AI is at most good for filling a page up with text that is remotely English and somewhat adjacent to the topic, but can barely handle one chapter worth of outline before it wanders off into the wilderness, and even then it's basically just expanding the outline into proper sentences; anything 'new' coming out of it appears to be errors or issues with prompting more than the model actually providing things 'on purpose' (being careful with the anthropomorphization here).
 
Last edited:
I documented with full disclosure my own brief experimentation with AI in "Wishes" (a spur of the freeuse branch). I ran it for a bit, posted a few lightly edited things for people to see (with commentary), then rewrote most of it using my original outlines I had fed the AI (and the occasional probably-accidental good thing the AI happened to spit out) and went back to writing manually. My basic conclusion was that, at least at the locally runnable level, AI is at most good for filling a page up with text that is remotely English and somewhat adjacent to the topic, but can barely handle one chapter worth of outline before it wanders off into the wilderness, and even then it's basically just expanding the outline into proper sentences; anything 'new' coming out of it appears to be errors or issues with prompting more than the model actually providing things 'on purpose' (being careful with the anthropomorphization here).

I normally use AI for planning and editing, which i find it very helpful for. I've recently started my own experiment with actual AI written chapters. My thoughts are much the same as yours, at best it's a page of words vaguely matching some parts of what i asked for. And continuity is a big problem, it's forever mixing people up or just making stuff up.

I have found it more useful if broken down into very small parts. For example, I found a photo online of a house I thought my character may live in, then i asked AI to drop a description of it into a paragraph for a chapter I'd written. It actually did a pretty good job. The problem was, it took so much effort to wrangle the paragraph out of it that I could have written it twice over in half the time. I'm persevering with my experiment though, I'm finding it oddly interesting as a distraction from normal writing.
 
Back
Top