--- title: AI Slop and AI-sounding Prose date: 2026-09-10 category: tech summary: Slop is empty AI filler. Removing the AI taste means editing and cutting fluff; in code, it means refusing over-engineering. draft: false --- A new word has been trending lately: slop — garbage text produced by AI. Most of the text AI produces is defective. The internet has run plenty of analyses, and plenty of de-AI skills or methods have been made. AI likes to pile it on with masses of ornamental vocabulary, and it also likes the “not X, but Y” sentence pattern. Honestly, seeing articles like this gives me PTSD too. AI-produced articles come out very empty. One especially big reason: when you give it input, you only give a title, not any concrete content. To break that feeling of emptiness, the AI can only add decorative words of its own. Even when you do give it concrete content, AI will still find ways to add more transitions and declarative sentences — this makes the reading experience very poor. Some people say that of all the internal documents at their company, 80% are AI-produced. No one will want to read these documents; they just throw them to AI and let AI do the analysis. Since the writers use AI and the readers use AI, these documents actually no longer have any real reason to exist. Such documents deserve to be called Slop, whether judged by a human or by AI. Because the junk words are so plentiful, they waste context space — and in engineering that is a cardinal sin. Here is another viral post: “What I can’t stand most about GPT 5.6: I ask it to make tomato-and-egg stir-fry, and it adds Dongpo pork. I ask if the pork is necessary. It says I’m right and removes it. I say fine, open a PR. The PR title is ‘Tomato-and-egg (no Dongpo pork)’ with a long comment explaining why this dish doesn’t need Dongpo pork.” That phenomenon is very typical in AI coding today. It is essentially the combined result of a large language model’s attention mechanism, its Prompt-completion logic, and its context habits. AI tends to add all kinds of impressive-looking but utterly unnecessary feature modules, exception handling, or helper logic into a simple function. This skill on GitHub exists to solve the problem of AI over-engineering. Anyone who pays attention to AI coding should have come across this skill: its goal is to make AI act like a lazy programmer and refuse any over-engineering. As for why “no Dongpo pork” shows up, that too has to do with the model’s underlying principles. AI models have no true forgetting/deletion mechanism, only an appending one. The model’s attention mechanism determines that once you emphasize Dongpo pork in the session’s context, that word’s weight gets raised, which then affects the model’s output — so “no Dongpo pork” ends up in the title afterwards. There are two ways to fix this: one is to start a new conversation, the other is to emphasize the output specification in the prompt, keeping only useful information in the output. I don’t reject AI-produced articles, but don’t show me AI slop either. Finding useful information in a big pile of useless information is a very tiring job. Actually, back in 2022, when I first used GPT-3 to write a paper, I already noticed the serious AI-taste problem: a paper that keeps reusing the same sentence pattern is glaringly AI. My solution then was human proofreading — first make sure the content is correct, then manually revise the sentence style, rewriting a long sentence full of ornamentation and transitions myself. That cut the obvious AI taste, but after reading for a long time you would still discover that this paper was actually an AI product. The problem may lie in the expression logic and the paragraphing — but at least it no longer read as AI at a glance. Even today, de-AI’ing is still that same playbook; it is just that some operations like “style plagiarism” or “role immersion” have been added. No matter what method you use, as long as it guarantees that a human can read the article normally, that is enough.