#coffeegoldfish

AI Generation and AI Assisting

First, I would like to point out that many AI checkers are unreliable, and just because something passes the AI tracker check doesn't mean that it hasn't been generated. For the purpose of this memo, I do not intend to make use of AI trackers, as I have used them prior to making this post and found them to be lacking. This reddit post goes more in-depth on how AI checkers are basically glorified pattern detectors and ways in which to fool them; additionally, OP provides a pretty good writeup of their own personal accuracy metrics, and what they feel are effective uses of AI checkers.

Another thing to add is that I am not considering AI spellcheck or grammar check as "use of AI" because I have noticed that Google had the temerity to start training their spell check using it, so now I have to write in the damn notes app. Anecdotal evidence here, as well as Google's own page on this and an eight-minute video from someone who explains all of the new features.

In my own anecdotal experience as a tutor, undergraduate, and graduate student, many people have gotten pretty good at going back and touching up to include just enough variation that large chunks are not easily recognizable as AI at first glance. A common practice is to write the general gist of an idea and feed it to AI to "clean it up" or "touch it up", after which you will likely edit afterwards to make sure it has enough of "your" voice to retain plausible deniability. Another way of using AI is to accessorize chunks of prose and fill in the rest with what you think happens in between. I'm sure there are other ways to use it, but the main flaw when doing this FOR FUCKING YAOI is that the AI itself has been trained such that it has one defined and easily-flaggable personality and dialogue style for Will, as well as one for Mack, and composes very similar speech and narrative patterns in fanfics from completely different users. A third identifier is that AI LOVES em dashes and uses them as much as possible - this is a fairly recent development, and while entertaining, it gets old fast for those of us who actually like to use the em (and en) dashes. There are other identifiers, and of course I'm only one person, so I accept that there is always a risk of error. Maybe you just write like that, or you read too much of the same fic and decided to replicate patterns for clout or notes. Who knows. With that being said, please see below.

The Writers

ALLTIMECHARLO

alltimecharlo's AI usage style is characterized by extremely frequent usage of the same words, phrases and stylistic choices in dialogue and narrative, consistent use of negatives in exact patterns, and a dialogue pattern I will be referring to as "not-not-just": "Not [noun/adjective]. Not [second modifier]. Just [half-sentence with other modifiers]."

A spin on this is the "not-just" or "no-no-just" pattern. See:

There is also "doesn't-just", "maybe-just", and "and-just-because":

Works cited are tape review, anywhere you want me, birthday boy, second serve, breathe me in, a soul in tune, and need a big boy (mattgabe).

DALSY

dalsy's AI usage is characterized by similar patterns to alltimecharlo's - a writing style that seems to be regular across works, but reveals similar pattern usage. The most damning of these is what I will be referring to as the "too-too" pattern, and its offshoots: "Too [noun/adjective], too [second noun/adjective] (sometimes paired with a sentence instead)."

A variation on this is frequently found with a singular "too":

As well as the "so-so" pattern:

Works cited are cross your heart, picture perfect, sounds like trouble, dumb about you, slipping through my fingers (nicojack), and heatstroke.

MAXENCEANDREBISSET

the evil was defeated before I was able to catch any of their fics. Sorry guys!

COFFEEGOLDFISH

This one is a special case - apologies for the extremely crunchy screenshot, but this author was part of a Twitter writers' group chat where she posted her wips and there was literally still formatting from the copy/paste and generators she used included in one of the documents. Additionally, she made an AI-generated tweet several weeks ago.

Conclusion