With this sudden amount of conversation of AI-assisted fics and such, is there a specific fic or fics that people are talking about as being "super popular" but obviously AI written? I feel so out of the loop
(Feel free to send me this on anon if you don't want to comment something publicly. I won't publish anything that names an author / fic)
I believe the uptick in conversation is in reaction to this. The Reddit post links to a conversation/document that explains a specific html markers which proves that a piece of text has been copied + pasted directly from Claude AI, and then includes a rundown of how many of these markers were found in a subset of very popular fics in the Heated Rivalry tag on AO3, some of which are extremely popular and successful (as in, very high stats, are talked about and recommended a lot, authors have name recognition).
This marker only exists if text was copied directly from Claude AI into the AO3 text box. It wasn't originally identified in a fandom context and doesn't occur if the text was copied from Claude into ie a Word doc or if a different LLM was used. It also doesn't differentiate between text that was completely generated by Claude and text that was only edited by Claude for formatting or spelling.
The main takeaway from this experiment is, in my opinion, that AI use in fanfiction is widespread and ubiquitous and not detectable to most people. There is no reliable way to tell if a text was written using AI because of how LLMs work and what they were trained on, but a lot of people ARE using AI to write fanfiction and the results are good enough to make AI use rewarding.
The other thing is that it complicates the way fandom as a community talks about AI use in fic. Most people don't want to read fic that uses AI, which disincentivises writers from disclosing their AI usage. People have been using AI accusations to bully and harass authors, which is bad because bullying is bad, but also apparently a lot of authors are using AI and vehemently denying it when asked. Reconciling all of this in a way that satisfies all the different interests and ideals about fandom conversations is going to be really, really difficult and probably get really, really ugly.
It also doesn't differentiate between text that was completely generated by Claude and text that was only edited by Claude for formatting or spelling.
Just to note that this can't be differentiated because LLM models (unless paired with additional methods, which Claude may or may be) don't track input in this way. All output is identical: it's statistically generated text, based on the input. There's no mathematical distinction between "editing" and "generating" for the model, though people try to draw a distinction when they talk about LLM use by humans.
Otherwise, yes, agreed. Presumably there's plenty of authors copying into word processors and *then* publishing to AO3, which would hide these artifacts.
I'll just say this: my mathematical research is adjacent to LLMs. I understand very well how they work. I do not think the mathematics of it is somehow evil. And I refuse to sit around and draw lines about "acceptable" use: is it ok to use LLMs to copy edit? What about to translate? Or to generate a list of names for background characters? To format?
You could spend a lot of time building an ethical framework to try and answer these questions in a coherent way. But for me it comes back to the fact that I fucking hate "AI" and the companies that peddle it. I do think they're evil. And I don't want to read anything touched by "AI" because I don't want to engage with their products.
But there's never going to be a way to know for sure, short of artifacts (that will very quickly disappear, now, due to the new knowledge of how they're generated). It's easy to just feel hopeless.
But maybe OpenAI is going bankrupt, so let's hope for better days. Or I don't know. Going back to just reading fic by friends, via mailing lists, maybe.
