The Heart of the Slop

Slop. It's a word you hear all over the place today, especially in conjunction with AI, but what is "slop" precisely? People throw it around rather loosely to be honest, so part of the problem is it gets used as a generic term for "crappy AI output," but let's talk about the actual mechanism that the word is supposed to be targeting, because when you understand how AI functions at a deeper level, the slop that sometimes comes out actually makes perfect sense. Think of it like the head on a really good beer that you need to shave off the top of the glass with a butter knife. You can't take an unimpeded sip of the quality stuff till you push the tasteless fluff out of the way and into the sink.

Many of you already have a sense of what slop is, even if you can't articulate the mechanism. You just know it when you see it. It lacks heart and substance. Words with nothing beneath them are as unsettling as a smiling face with nothing behind the eyes.

To understand AI slop, you first have to understand what's under the hood of chatbots like ChatGPT and Claude: Large Language Models (LLMs). LLMs don't think. They just predict what word might come next given a context, then the next, then the next.

Here's a simpler way to see it. Finish this sentence: "Happy birthday to ___."

Chances are your brain filled in "you" so automatically that you heard "you" in your head before you registered the blank. Your brain did what LLMs do: predicted the most likely next word in the sentence.

An LLM is that same trick at an enormous scale. The model has absorbed a staggering amount of human writing, and it's gotten so good at predicting what comes next that it can keep going for paragraphs. The result looks like a conversation, but it's just a string of words constructed based on a context - the prompt.

Fun experiment for you: ask your favorite AI chatbot to summarize a book that doesn't exist. It will quickly give you a plot summary, probably formulaic, and different every time you ask. Since there is no book to actually summarize, the context is wide open, and the more open the context, the more possible next words there are at every step. With no source or structure to anchor it, the output drifts.

Next experiment: ask your chatbot why apples are delicious - or pick a food you really love. It will string together sentences on the topic, and some of them will be good points, since the internet is full of apple praise. But I'm betting you could explain your love of apples better. That's because you have something to express: a feeling, a stance, a memory of a really good apple. Your brain does sequence words much the same as an LLM, but the words are an expression of your experience and feelings. The LLM is just the words.

Strangely, when you read those words, you might feel something behind them anyway. That's the power of the illusion. It just doesn't happen reliably. Some of what comes out of the tap is quality beer, and some is just tasteless foam you need to shave off the top. The key thing to understand is that the LLM is not thinking or feeling, and if you do sense substance behind the chatbot's words, realize that it is your own thoughts and feelings being reflected back at you by a very sophisticated mirror.

Maybe even more surprising is that the term "slop" existed well before the current AI revolution. Humans are perfectly capable of producing great quantities of slop. I recall a popular service for businesses looking to rank higher on Google - they'd hire people to write blog articles stuffed with specific keywords. The business handed over a list, and the writer worked every term in. The articles rarely made any sense, but that wasn't the point. It was slop written for a search engine's crawler, not for people. We were doing just fine making soulless slop without ChatGPT or Claude.

Words have always worked like a mirror. When something moves you, part of what you’re seeing is yourself. When a human writes it, there is someone on the other side, and you connect with them, sometimes across an ocean of time. Connection is arguably the point of all language. You might view AI as a simulation of that connection without the need for an actual human on the other end, and that’s not wrong. But a more constructive frame is that you are holding up a mirror and honestly examining yourself. The difference between quality and slop comes down to whether you found anything of value in the illusion.