Mr Schachtel has a great little musing out on AI writing #1536470, #1536926 #1549671 (originally at American Greatness, an outlet I've never heard of)

"AI has many incredible applications, but AI writing is mediocre by design. It’s fine. Not great. Just fine."

It’s suitable for the vast majority of the population, especially since most people currently can’t identify the difference between AI writing, AI-assisted writing, and purely human, unassisted writing. Undoubtedly, there’s a place for all of these calibrations.

Mmm.. maybe. But with infinite access and information overload we gotta filter/sort somehow. And the writing doesn't label itself: green for snarky den stuff, red for boring uninteresting AI slop.

Ergo filtering and curating: #1554013

AI is trained to be a teacher’s pet, one who calculates what the “teacher” (AI model trainer) prefers and decides to write that way to achieve the reward. AI writing is habitually mediocre because of this.

I dunno, I think you can tell them to be spicy. Exciting. Dangerous.

More and more content online is now written entirely by AI, so AI models are now training on AI writing at an increasing rate, which likely means that this mediocre writing problem will get worse before innovation arrives to make it better.

"You can’t tell chatbots apart by their prose because it’s one big circular learning environment"

if you tell an AI to “write an article” about something, it does not have an independent perspective, so it fills the space with pieces of information that have been spliced together from a variety of places.
Due to the aforementioned institutional hurdles to creativity, it’s become clear to me that great human writing (with or without AI assistance) is needed now more than ever.

ACTUALLY:

You can direct it more granularly than that: give it 3-5 components (often standalone articles, or in my case, previous SN rants) and ask it to write in the voice of yourself (again, in my case, trained on my ~500 whatever published articles online).

In recent months I've asked Chat and Grok to make some initial drafts for me to support some of my article writing. It's been pretty useful: especially when I haven't quite figured out a story yet but have the semblance of a piece, with some specific components -— often anchored in some previous writing or, increasingly, some in-the-moment musings.

What I've noticed recently is that it's exceptionally good at creating succinct, high-level summary of wider arguments (this is also Chat's major drawback: too vague, too hedging, too nonspecific). And the first few sentences from draft articles, both at Grok and Chat are impeccable; fucking fantastic. Usually direct copy-paste into my draft document.

...then it repeats itself, doesn't make progress in the piece, starting making iconic AI slop such as one-line paragraphs.

So there's still some human writer oversight necessary.

The second thing it's really really good at is polishing, and it enters the sausage production toward the end of the process. I often upload my almost-finished drafts and ask it to correct, tighten, devil's advocate, identify the weakest point etc.

It will, sycophantically, provide endless suggestions here and I don't have to listen to it/agree with it, and can obviously stop at any point But the signal isn't zero and the effort is pretty low, so it's usually worth doing.

Also, titles. LLMs are fucking excellent at generating catchy titles; ask them to give a hundred different ones and it'll vary in tons of good ways for you to choose from.

Anyway, I used to be way more negative on LLMs producing viable drafts. I reluctantly confess that they have some use in the production process... at the very least to pull together some detached/unstructured thoughts and make them into a first draft you can work with.