I’m going to try and describe everything that follows tastefully. I rarely describe things tastefully, and this is not a situation that lends itself to doing so. Let’s just say that you may not want to
Google any of the terms I bring up, and you definitely don’t want to do so at work.
Let’s begin.
In the spring of 2023, we were still getting used to generative AI. Generative AI, in turn, was still learning by scooping up vast amounts of information from the open internet, and people were developing ways to discover its sources by prompting models with specific styles, terms, or tropes.
Every group has jargon, from “red teaming” and “MFA” in security to Amigurumi and “frogging” among knitters and even that little turn of phrase you and your best friend use because of one college roadtrip. When AI shows that terminology, especially in the right context, it shows where it got its material as clearly as saying “fries” or “chips” shows where you grew up. The smaller the subculture, the more obvious it is when AI has picked up a word or concept from it.
Fanfiction isn’t a huge subculture to begin with. To post it, you have to be a fairly intense fan of a particular work, willing to write your own stories based on that media, and finish enough of it to put online. The average person probably won’t be familiar with terms like “one true pairing,” “dead dove,” or “slashfic,” and the range of sites where an LLM can pick them up is pretty small.
And then there’s omegaverse.

The wolf at the door, or whatever
Like a bizarre number of phenomena in the modern world, the terminology associated with omegaverse began with a deeply flawed (and since very redacted) study of wolf behavior. You’ve probably heard of it if you’ve spent much time around the most tedious men online: the concept of “the alpha,” the Born Leader Manliest Man Who Ever Manned, bursting with testosterone and at least 60% Axe Body Spray.
In reality, “alpha wolves” are just the fathers of a wolf family, and the mothers have equal status. Just wanted to put that out there.
“Alpha,” “beta,” and other terms have propagated in a variety of…fascinating…ways since people decided that the original study applied to humans. We’re not here to discuss werewolf romance in general, though, nor (thank God) the sweaty fever dreams of the willfully inadequate, with their “sigmas” and “soyboys” and whatnot. No, we’re going to talk about the omegaverse, which took a whole different angle on the subject.
It’s important to keep the general concept in mind, though, because “alpha” itself wouldn’t necessarily serve as an LLM fingerprint. We have to go deeper.
The omegaverse is sort of science fiction or fantasy, sort of romance. It divides humanity into three types: the alphas (not necessarily male, though usually), the betas (any genders, normal people), and the omegas (innately submissive in some ways, any genders). It also gives alphas and omegas some of the traits associated with canine relationships, like pheromonal reactions and the romanticized “mating for life” of wolves.
The setting gets very complicated and very anatomical. A glossary of the omegaverse involves words either completely exclusive to it or those that generally don’t otherwise appear in connection with humans. When an LLM generates text about one lover “marking” another as a form of commitment and it’s not just getting matching tattoos, or uses terms like “knotting” and “heat” outside of a zoological context, it pretty obviously has scraped data from a particular few fanfiction sites.
In that far-away spring of 2023, a fanfiction writer discovered ChatGPT was doing that very thing. Plugging an omegaverse prompt into Sudowrite, they got back a passage showing that the LLM knew the territory as well as an alpha knows their omega’s…well, never mind.
This did not go over well.

Strong feelings? On the internet?
Fanfic writers vary in age, race, gender, sexuality, and a lot of other demographics, but they generally have two things in common. First, they’re very online: the hobby may have started with mimeographed newsletters about Star Trek, but it’s kept pace quickly with technology, and practitioners are generally familiar with the latest developments, AI included. As with almost everyone who talks about AI, many writers have very strong opinions about it.
Moreover, fanfic and monetization have a fraught relationship. The vast majority of fanfic can’t earn its creators any money: that’s how it falls into the fair use loophole in copyright law. Cases where writers skirt the edges of that principle, whether they involve taking commissions for particular stories, selling bound and printed copies, or Amazon’s failed Kindle Worlds, have been highly controversial within fandom. Some say monetization goes against fanfic culture; many others are worried about the legal problems it could create.
If insiders earning money from fanfic causes drama, an outsider doing so is naturally going to raise hackles, especially when it’s an LLM. When fanfic authors found out what ChatGPT was doing, they got upset. Then they acted.
Within two days, participants in the “Knot in My Name” campaign had written 450,000 words of new omegaverse fanfic. The goal was a kind of crowdsourced data poisoning, the aim to make AI-generated text default to tropes, terms, and other devices that only a minority of a minority of its audience would be looking for.
It’s a clever idea. It didn’t really work.

Protest, progress, and revelation
While 450,000 words is impressive on a human scale, it doesn’t make much of a dent in GPT-3’s source material. Furthermore, LLMs had already started using reinforcement learning from human feedback and other filters in 2023--after all, weird quasi-canine sexual fantasies are far from the worst material on the internet. It’s possible that Knot in My Name and the omegaverse in general provided guidance or impetus for refining those techniques, but the companies who make major LLMs have never said so.
On the other hand, raising awareness is itself a major goal of activism, and the omegaverse situation definitely did that. Many fanfic writers responded by locking down their accounts or posting that they didn’t give AI permission to use their works; while models like ChatGPT or Gemini may not pay much attention to such statements, it’s an indication of public sentiment that could help shape company action or regulatory policy in the future.
I would also bet my not-very-significant savings account that more people knew how LLMs get their data after Knot in My Name and the associated publicity than did before. More people probably opened articles about it than would have even glanced at general information about AI data. Even when we’re dealing with the newest technology, the oldest principle applies: sex sells.
In this case, it also illuminates, going not just beyond the omegaverse but beyond fanfic itself. Discussions of knotting and alphas made Sudowrite’s data source obvious. If GPT-3 scraped such a specific niche of the online world, though, it certainly didn’t leave the rest alone. Your comments on a newspaper article, your blog talking about your grandmother’s recipes, and the emotional post you made after a major life event are all grist for the mill.
Maybe it just took a story about cute guys in heat to make some of us realize that.
Sources
References cited in this piece. Last verified on the published or revision date.
- 01
- 02
- 03
- 04
- 05
- 06