LLM Writing and Me

🇺🇸USEnglish

LLM Writing and MeCasual

This blog hasn't used LLMs to write drafts or finished pieces, at least not yet (though it's not entirely LLM-free; I do use LLMs for proofreading things like spelling and for parts of the code). That said, browsing around lately, LLM-written content has become far more common.

The prose styles have grown noticeably similar, and on top of that, a lot of the sentence constructions feel un-Korean in their phrasing. Most tellingly, the writing itself lines up in what you'd call high information density, text packed so tight there's no room to breathe.

These days, no matter what site you visit, you're likely to run into writing that looks like it came from an LLM. The style is recognizable: a short introduction, a problem statement, several subheadings carving up the argument, a fair-handed summary of pros and cons on both sides, and then a conclusion that's just a little too tidy. Not every piece follows this pattern, and there's nothing inherently wrong with the style itself. If anything, it tends to be easier to read and more informative than the average piece dashed off by a person in a hurry.

In Korean web novel communities, there's a term called brick-like text (벽돌체) . It originally referred to writing with long, dense paragraphs that look like a wall of text on screen. I'd argue that LLM-generated writing exhibits its own kind of brick quality. In this semantic density, meaning flows on without a moment to breathe. Every paragraph has something to explain, every sentence has a role to play, and so even though the structure is neatly divided, what you're left with after reading feels like one solid mass: semantic brick-like text.

Human writing inherently contains resting points, what you might call unnecessary ornamentation (군더더기). To stick with the brick metaphor, there are holes in the wall, and where a writer places those breathing moments is itself a marker of their individual voice. There are parts a writer wants to describe even when they don't serve the central topic. In informational writing, this often shows up as a slight detour into the author's own specialty, something adjacent to the point being made. LLM writing has none of that. It's information-rich, clean, and uninterrupted.

And that is both its strength and its weakness. There's too much information, packed too tightly, and it's immaculate. The expression is polished, neat. As a result, unlike human writing, there's very little that needs to be stripped away. Whether this is a matter of token optimization or simply efficiency is hard to say.<br>Compared to an equivalent passage written by a human, LLM writing holds more information in fewer sentences. You notice this as you read: not much text, but a lot of content packed inside it.<br>One distinctive feature of LLM writing, if you can call it that, is that it tends to use a lot of full stops. Human writing, by contrast, leans on commas, or runs sentences together into fragments.

This is especially noticeable in Korean, my own language. The Korean people actually speak runs on connective endings (연결어미), letting sentences stretch out and leaving logical relationships deliberately vague, but models break those chains and make the connective relationships explicit. They can't tolerate the ambiguity.<br>The thing is, people sometimes leave that ambiguity in on purpose. The shape of the sentence itself is communicating that the speaker isn't certain. Though I should note this is just my hypothesis; I haven't verified it with data.

Certain expressions also keep cropping up. The problem is that every information site is now plastered with them, and you can even see people starting to imitate the style. Being a programmer, I tend to use a lot of overseas sites and information aggregators by the nature of the work, and nearly all of them are saturated with LLM prose. The most frequent pattern I keep seeing is this:

Rather than dismissing a technology or methodology outright, LLMs often frame things as a trade-off where the advantages of A diminish as the disadvantages of B grow, keeping a structural retreat path built into the sentence itself.<br>Common sentences from Korean LLMs include things like "The cost doesn't disappear. It only shifts to XX" , along with a general preference for neutral phrasing that avoids hard negatives.

On top of that, there's the sterile/aseptic speech (무균식 화법) pattern, things like "While it may be useful in certain situations, limitations exist in other respects." This makes the logic airtight. The writing becomes safe, impossible to attack. Sentences that pre-emptively account for edge cases and hedge against them are defensible as information delivery, but they make it very hard to infer anything about the speaker's personality from the text.

Some people might disagree with all this, and AI fiction writing is probably where the pushback is strongest. People say story-driven AI output feels disconnected and...