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

Author’s Note: This article makes use of generative AI for examples of AI-generated content only.

Over the past few years there has been a complete and total shift on the internet. The fundamentals of online interaction have changed entirely. Online content has taken on a new, almost unrecognizable form. Thanks to AI, it is now easier to rapidly mass-produce content than it has ever been before. Anyone can become an influencer now, anyone can write a video essay. The possibilities of content generation are endless. However, this begs the question: is the quality of this content good?

No, it’s not.

The proliferation of AI-generated content has literally stagnated the entire Internet. Everyone talks the same now. Every infographic and poster and flyer looks the same. The content doesn’t even look good. It’s all stale, self-important, overindulgent nonsense that is so obviously generated by a computer that it’s frankly pathetic. People don’t even take the time to check their AI-generated content, yielding horrendous writing and ugly images full of text and elements to the point where you can barely read anything.

Given this meteoric and exponential rise in the amount of AI-generated content, I believe it to be necessary that I show people how to spot it. Note that I am far from unqualified. My mom moderates a Facebook group for pencil drawing that has over 400,000 members from around the world, and a strict, zero-tolerance no-AI rule. Whenever she is suspicious of a post being AI-generated, she sends it to me, and I am generally good at picking it out—sometimes even down to which AI model generated it. I know how AI writes, and I know the hallmarks of AI-generated images—two of the most common forms of AI slop on the Internet. And that is what I am going to teach you, dear reader—how to spot it. 

Full disclosure: as a form of doing research for this guide, I contributed 20 whole dollars towards human extinction in the form of a one-month ChatGPT Plus subscription and utilized it for research and casual stupidity only. I am opposed to using it in place of my work. The free tier limits are hilariously low, especially for the research I’m conducting, and you don’t really have access to advanced models on that tier. So, much of the actual AI-generated material will be generated by yours truly. Yes, you can absolutely question the ethics on this. I sure am. I’m not getting that $20 back. 

Know Your Enemy: The Mundane Technical Side

Author’s Note: If you’re not interested in all the tech stuff, feel free to scroll down for the signs of AI spotting.

While everyone knows about AI and AI slop because of its ubiquitousness on the web, not everyone knows exactly how it works. I do, because I believe it’s important that you know your enemy inside and out. That’s why I spent money on it. So, before we “delve” (an AI-favorite word!) into this tutorial, let’s go over a few concepts on the more technical side.

The first term I want to go over is “tokens” or “credits.” Most text-based AI models (Large Language Models or LLMs) will prefer the term “tokens,” but they are pretty much the same thing. Tokens, however, are specifically text. It is not a word or a letter but rather a small grouping of letters, generally four letters. Take the word “jumps”. It is worth two tokens. The first token is “jump”, and the second token is “s.” AI services generally charge by the token—typically a fraction of a fraction of a cent. OpenAI’s ChatGPT charges approximately $10 per 1 million tokens, which brings their per-token rate to $0.00001. 

Credits, meanwhile, refers to the abstract “currency” used to pay for those tokens. The more tokens and processing power you use, the more credits you consume, and the more you pay. This also applies to any sort of computing for AI—image, video, audio, and any other sorts of generation. When using an online AI image/video generator you are likely to see the term credits being used. On those sites, you can purchase them outright or get a subscription that entitles you to a certain amount of credits per month.

Discussing AI billing is where we get into the idea of the API. An API, simply, is a way for computer programs to use other computer programs. This is not just for AI programs—there are also APIs for Spotify, Google Maps, Twitter (X), and pretty much every other website on the Internet, for any amount of purposes. A developer or user makes an API request to a program, and the program returns whatever information or generation they seek. When a program provides an API, they typically don’t want users to have unlimited access, as it would strain their servers. So, they give each user an API key which is included with the request and completely tied to their account. 

A practical explanation, in the words of user SurprisedPotato on Reddit: When you send an API request, the program effectively queries the server like this: “I am (XYZ user) [key]. Find all posts within the past 3 months featuring the keywords ‘banana for scale’ [request].” The system returns the request and bills you appropriately for how much resources the request consumed.

Now, where does billing come into play, and how does it tie into credits and tokens? Well, like I said before, AI services generally charge by the token (or million tokens). They’ll eat the bill for non-paying users, at the cost of severely limiting their usage to a couple of queries per hour or day if at all. Your ordinary heavy user will probably just pay a flat monthly subscription fee, generally $20 or $200, for enhanced usage. However, APIs are different. They are, theoretically and for all intents and purposes, unlimited. The limit comes in the form of billing. The more tokens and queries a user makes, the higher their costs. Some users spend thousands of dollars a month on APIs, especially when it comes to having AI code for them (a.k.a. “vibe coding”) or building “vibe coded” websites. That’s typically because those tasks are incredibly resource-intensive and consume a lot of computing power. The users who do this typically use the flagship advanced models (such as Claude Opus or ChatGPT Astra) because they’re really powerful. Hope they’re working on something that’ll pay for it, or else it’s useless money and compute spent…

But now that I’ve gotten into the technical side of AI, let’s get into how to spot when someone’s using it. This first part deals with text, and the next part will deal with images.

Cheap Models: Short. Snappy. Insufferable.

I’m sure you’ve all been scrolling on Instagram Reels or YouTube Shorts (no judgment here…) and seen some video about an interesting topic or place. For example, the largest Walmart in America. The video probably features a Google Earth flyover of the given topic, spinning in a circle around it. And the script probably sounds something like this:

“Most people don’t know this, but the largest Walmart in America is so massive that it has an escalator inside. It’s not in New York City, it’s not in Los Angeles…” and then it goes on for 3 minutes in a circle before finally revealing that it’s in Albany, New York. (This is true!)

Or, you’ve been on Twitter (X) and gone to the Chart Data or Pop Base pages, or really, any popular Twitter (X) account. There will probably be an interaction like this…

Chart Data: “Today in 2007, Kanye West's "Stronger" reached #1 on the Hot 100. It became his first Diamond single.”

Top reply: “An iconic record. "Stronger" reached #1 on this day in 2007 and still holds up as one of Kanye's most important hits🔥”

Neither of these things sound like a normal human wrote them. The Walmart video doesn’t sound like a typical video essayist (it lacks pretentiousness—and I mean this as a compliment to video essays) and the Chart Data reply feels uncanny and adds absolutely nothing to the conversation. There is a reason for that: cheap AI models. Generally speaking, “content farms” need to pump out content as quickly and cheaply as possible to maximize engagement. The same logic applies to these bot accounts on social media. It all comes down to money from engagement—the more views you get, the more money Twitter (X) pays you, which leads to stuff like this. The owner of the bot farm prompts their models and APIs to scroll social media and respond in a certain way. I’ve seen bots who hate everything and are frankly mean, bots who are super cheerful and positive but add nothing to the post, and bots that make risque puns with every third word trying to sell you their AI-generated OnlyFans. I love the Internet…

So, how do you maximize output quantity while keeping your API costs down? Simple: Use a cheap model. Most AI platforms will offer multiple levels of “thinking”—a low one, a standard one, and a high one—with price increases depending on how much “thinking” is done.

The typical “syntax” of these cheaper, more efficient models is extraordinarily predictable. 

Short sentences. A few words. And the rule of threes.

Each sentence gets its own paragraph, like this.

And it’s always a mic drop moment. Or it tries to be.

Did those sentences make the hairs on the back of your neck stand up? It seems awfully familiar, no? If you spend… any amount of time, really, on the Internet, you’ve probably seen a post that’s written in that exact style—not word for word, of course, but still exceptionally similar. Short, snappy sentences that try to be poetic or mic-drop-y. To the point where I can only describe it as insufferable.

We all know the “It’s not X, it’s Y” giveaway—it’s perhaps the most common hallmark of AI-generated text. Think “This isn’t just an idea—it’s a movement.” Or the title of this article (I came up with that myself). But do we all know about its younger, more annoying brother? That younger brother goes by the name “No X. No Y. Just Z.” It has several derivatives, but the general flow is the same. AI loves the rule of threes. The rule of threes is pretty self-explanatory, but if you don’t know, it’s the concept that three items in a list is more effective and memorable than other numbers. That used to make sense, but now it’s a giveaway that someone’s probably using AI.

There is a particular “favorite” example of AI usage that I have witnessed. Reader, have you ever been to Chili’s, Red Robin, or Applebee’s? I hope so, because you’ll have needed to in order to understand what I’m talking about. Now, if you have, you probably recognize those little tablets on every table. You’ve probably seen a six-year-old tapping away on one of those games. I, at my grown age of twenty-one, still enjoy those things. I have a long streak of ZTrivia victories (I think I have about 15 consecutive wins in the past few times I’ve gone to those restaurants) and excel in geography, history, and music. I always use a cat avatar and the screen name “Fancy Creature.”

But enough about my trivia record and how I enjoy playing kids’ trivia games at restaurants. When playing ZTrivia in the past year or so, I’ve noticed something interesting. After most questions—about three-quarters of them—it’ll present you with a fun fact in the most enthusiastic tone possible. “Did you know that XYZ was ABC? Talk about a funny coincidence!” is basically the typical phrasing. 

If the writing wasn’t a giveaway enough, I’ve also noticed this weird thing that any normal and sane person would not notice, but I am not normal, nor sane. Anyway, the text in the game is not capable of rendering special characters, such as… the em-dash. As a result, it leaves nothing there, and a glaring typo where two words are mashed together. Take this sentence as an example: it's missing an em-dash. This is entirely a suspicion on my part, and I am happy to be proven wrong, but my theory is that the developers of the game asked an LLM for a bunch of questions and answers, as well as fun facts, and told it to write it in a fun and friendly style.

Author’s Note: As I’m sure you’ve seen, I have a lot to say on this topic. There’s a certain limit to the length of articles, and needless to say I have surpassed it. So, go check out Part One-B of this article for more: we’ll be getting into the more advanced AI models, as well as some universal quirks you’ll see with LLMs.