“This Isn’t Just a Tutorial—It’s a Survival Guide”: How to Spot AI Slop, Part One: AI-Generated Text (Part One-B)

This is Part One-B of my article about AI-generated text. Part One-A dealt with the technical side and less-advanced LLMs. So, if you haven’t read Part One-A yet, please go read it, you’ll be missing important context for this second half.

Expensive Models: Long-Winded Word Salad, Weird Metaphors, and the “Shy Girl” Situation

As opposed to “instant thinking” or “standard thinking”, many AI models also have an “expert” or “high thinking” level. This means that it’ll spend more “tokens” and generate longer, more in-depth—and theoretically, more intelligent—responses. Amateur vibe coders use it. Writers do too. And this is where the “Shy Girl” incident comes in.

If you have approximately three hours(!) of free time, I highly recommend Frankie’s Shelf’s YouTube video on the book “Shy Girl.” But if you don’t, here’s the lowdown:

In February of 2025, author Mia Ballard self-published a book called Shy Girl. It did surprisingly well at first—officially published by Hachette in the UK that November with plans to release it in the States this year. But in early 2026, people began to notice something was up with the book. There were too many nonsensical metaphors, drawn-out sentences that meant absolutely nothing, repetition, specific adjectives, and other linguistic quirks that made eagle-eyed readers take an even closer look.

In January 2026, Frankie’s Shelf uploaded the aforementioned YouTube video where he reads the book and provides evidence as to why he thinks it is AI slop. That March, the New York Times provided evidence to the publisher, Hachette, that indicated approximately 78% of the work was AI-generated.

Ballard adamantly denied the claims. She said that she believed her editor had run the text through an LLM without her knowledge, and announced plans to take legal action. As of this article’s publication in October of 2026, nothing new has come out since.

From what I have seen, the “higher thinking” models either repeat the “snappy” sentence structure from the cheap models before, or write long-winded word salad paragraphs that mean absolutely nothing.

However, this “high thinking” model of AI isn’t really commonly seen on the Internet, but you should learn how to spot it anyway—especially if you’re an author. I took the time and my $20 subscription to ChatGPT that I’m determined to get my money’s worth out of—despite my moral objections, and asked it to write a story on maximum thinking level. If you’re wondering how much thinking it does, the story took approximately 10 minutes to write. You can access the full text here (saved in Pastebin for longevity), but I’ve collected some of the highlights that serve as dead giveaways.

• “[The island] sat seventeen miles from the mainland, far enough out that the sea felt less like scenery than government. It determined who arrived, who left, what things cost, whether the school opened, when bodies could be buried, and which promises survived long enough to matter.”

• “Pastor Deen saw nobody. He considered this significant. Mara did not.”

• “For one second the lantern room became transparent. Mara saw lines of light crossing the ocean. Not beams. Histories. Every window ever illuminated. Every match struck. Every face seen. Every moment reflected outward. An infinity of images racing away from Earth at the speed of light. Mara understood before Jonah explained.”

These are some more examples of AI “word salad.” Even within the context of the story, a lot of these words really mean nothing. They make sense grammatically, but they’re needlessly long.

Okay, so now you’ve seen some pretty good examples of the lesser-known AI tells. The more “thinking-intensive” AI models are better, especially at writing lots of text, but still experience some of the shortcomings that the cheaper ones do. Like the “bombastic” sentences or unnecessary paragraph breaks.

Universal Linguistic and Textual Quirks

There are certain linguistic “quirks” that, no matter how powerful of an AI someone uses, will pop up in writing. AI loves using specific words. Here are some of its favorite words, linguistic quirks, and turns of phrase:

• Vibe(s): (“No X. No Y. Just Z vibes.”)

• “In a world of…”

• “And honestly? That’s iconic.” (or similar)

• Node (AI really likes this one)

• Enigmatic

• Just (pure/classic) X.

• Tapestry (abstract—“intricate tapestry of bla bla bla…”)

• Y rather than X

• “Serves as a reminder” rather than “is” (or similar) — AI really hates copulas and will draw out sentences as much as possible to sound “intellectual”

• Paragraphs of almost perfectly similar length (“square” paragraphs)

The final tell, in my experiences of AI detection, is that something is too perfect. Yes, human editors exist to make your text as good as possible. This article has been run through editors. But if someone you know suddenly changes their writing style into something quasi-poetic and self-indulgent, they’re probably using AI.

Repeated Elements in Stories and the “Elias Thorne Effect”

When I was doing research into how AI writes, I stumbled upon what is unofficially known as the Elias Thorne effect. If you’re not conscientiously opposed to using AI, go ask any AI to generate a story in exactly 10 sentences. I did it with five different models: ChatGPT, Claude, DeepSeek, Gemini, and the Google AI Overview. I came to several conclusions: it really likes lighthouse keepers (typically named Elias) who have been alone for 40 years, finding mysterious bright lights or shimmering metaphors, and women named Maya or Lena.

In other examples of me tooling around with it, there are other elements of stories that AI likes to use. I’ve had it try to come up with story and plot ideas—most of them are endless cycles of abstract nothingness or extremely derivative, no matter how creative you try and get with it. Another common character surname is “Chen”—especially tied to characters named Maya. When a character is Chinese, their name is probably Li Wei or Li Qiang. These are all statistically common names, but not extremely common—well, until AI started using them constantly.

My hypothesis for this phenomenon is that, since LLMs are probabilistic and predict the next most likely token, it wants to use common names. Not names that are too common or expected, like John or Jane. And it wants to sound cool. The names Maya and Lena and Elias are objectively cooler than names like John and Jane—and they share certain letters and tokens as well.

Below are the results of the fun little experiment I did with the five different AI models. The prompt is visible in every screenshot, so you can try it for yourself if you wish. Spot the patterns! All the screenshots were taken by me. 

ChatGPT:

Claude:

Google Gemini:

DeepSeek:

Google AI Overview:

Below is an incomplete list of things AI loves to disproportionately include in stories and story ideas, sourced from both my personal experience as well as users across the Internet:

• Lena

• Leo

• Elara

• Maya

• Lira/Lyra

• Mara

• Aria

• Elias

• Iris

• Marcus

• Silas

• Kael/Kaelan

• Marisol

• Lydia

For surnames:

• Chen (so commonly used that someone on Reddit said their AI story had so many Chens in it that it said “not related to all the other Chens”)

• Voss

• Vance

• Vex

• Li

• Thorne

• Blackwood

For place names:

• Eldoria (the most common one by far for fantasy worlds)

• Valoria (also for fantasy)

• Oakhaven

• Elmwood

• Whispering Woods/Hollow

• Willowbrook/Willow Creek

And finally, AI’s favorite cliches in storytelling:

• Quantum (were LLMs trained on the Marvel Cinematic Universe?)

• Fingers tracing something

• Pulsating crystals of power

• Gazes flickering

• Moral endings that wrap everything up

• “The world is holding its breath”

• Foreheads touching as a form of intimacy

• Things smell of two things, one of which is an abstract concept and isn’t a smell (e.g. “The room smells of mildew and despair.”)

• Inanimate objects having metaphors and personification (“The stone remembered the rain.”)

So, why does it do all of this?

There is, actually, an answer for this! Two, actually.

1. LLMs are trained on the Internet, especially Reddit because it’s effectively copyright-free. Your average Reddit post probably oozes of millennial-isms, which is exactly what AI is trying to replicate. At least ChatGPT isn’t drawing a fake mustache on its finger and putting it up to its lip, or using one of those fake beer apps on its phone. Yet.

2. When training AI, it will often go through a phase known as “Reinforcement Learning from Human Feedback” or RLHF. During RLHF, a human will prompt the model to do something for them, such as writing a short story. When it provides the answer, the human will score it based on how much they like the output. The model weights will note the score, and like how an animal is trained, will learn what gets high scores and what doesn’t. Apparently, AI researchers love these things that I’ve outlined before, and give the model high scores, which encourages the AI to do it more.

Next Step: Identifying AI Images

The second part of this article will deal with how to identify AI-generated images. I won’t be diving into how to spot AI videos—there are incredibly obvious giveaways if a video is AI. But people don’t often consider the minutiae of AI-generated content beyond the obvious ones. I do. So, look out for the second part of this article, to come out with the next Cyclone Chronicle refresh. In the meantime, here’s…

Some Further Reading and Watching (HIGHLY recommended!):

Wikipedia - Signs of AI Writing

Fractal Philosophy (YouTube) - How to Survive the Death of the Internet

Frankie’s Shelf (YouTube) - I'm Pretty Sure This Book is AI Slop

Hunting The Muse - How to Spot When Writing is AI

Daytona State University - AI-Generated Common Indicators

Leonardo Couy (Github) - Signs of AI Writing

Language Jones (YouTube) - How to Detect AI Slop