The temptation of LLMs as slop generators

I recently fell down this trap myself. I had an assignment i was rushing, and I suck at writing prose (just look at this post...). I decided to simply note down everything as bullet points. The thoughts were mine, and did research the luddite way of crawling through Google Scholar, Sci-Hub, and random blog posts. I liked the output very much - it conveyed all my arguments and had all the relevant facts. However, it was a bit disordered and probably didn't meet the mark scheme.

So why not put it into an LLM? Surely a language model would be good at language and simply transforming text from one form to another. At first glance, it all looked fine. I see the stats, I see my citations, and there are now full sentences.

But actually reading through the text from top to bottom, it was obvious how shit it was. Important points were missing, and unimportant points were emphasized. The style was disgusting, full of weird em-dash usage and corporate speak I'd never use myself. Every time I corrected it, another mistake would either pop up elsewhere, or it'd over correct to some hippy-speak. This was not a model-specific problem. No matter which model I used: Kimi K2, GPT-5.2, Claude Opus 4.5, Deekseek R2, Gemini 3 Flash, none of them could follow my specific preferences.

$10 and an unreasonable amount of time later, I finally gave up and started rewriting from scratch.

Nobody likes to read AI slop, so why do we keep making more of it?

It's not about the quality of the content, but the aesthetics of delivery.

Not everyone has English as their first language, and even for native speakers, it takes effort to construct prose in a way that flows well and sounds eloquent. In academics, despite linguistic skill not necessarily reflecting the quality of content, bias easily seeps in based on how things are worded.

For example:

• Even in software engineering where our assignments are usually to write code, we are assessed on reflective reports and essays where snaky wording can allow a team member with 0 commits to still score a 1st.
• In 2024, during a hackathon by Huawei, the team that placed first in the technical round completely fell out of the top 5 simply because English was their second language, whereas my team that scored 4th in the technical round got bumped up to 2nd due to presentation skills alone.

Especially when the substance is entirely beyond the ability of the assessor to comprehend, writing quality becomes the subject of judgment.

LLMs provide an easy way out... It is benchmaxxed on a very specific style reminiscent of how we were taught to do so in middle and high school English: Vary between sentence structures and lengths to avoid sounding repetitive, structure content with numbering to make it easy to keep track, and use bigger words to sound well educated.

However, using LLMs take the soul out of content. It has no understanding of what is important, what's tangential, and happily makes up details not specified. It does put intelligent sounding text though, and that's what the vast majority of non-technical people I've met actually prefer anyways. Instead of reading an email, business report, or anything of substantial length, managers, acquaintances, and even family I know would simply dump it into ChatGPT or Gemini and read their hallucination-ridden output instead.

Furthermore, using an LLM is like gambling. Maybe you'll one-shot it, but probably not. Maybe the next try will be the one where it gets everything right. The probabilistic nature of LLMs as well as the dopamine hit of feeling like you've been productive without actually doing anything. "Ah I've done what would usually take me the whole day to do. I can slack off now" is always at the back of my mind. It makes you feel an illusion of productivity and forget that most of the time spent writing is actually time spent thinking and refining.

Heck, put in "Write an essay on why LLM Slop writing is so tempting" to ChatGPT and it writes a better blog than I do with all my points and more.