"Are you expecting that [writing with LLMs] eventually will be much easier?"
It's weird, because it seems like it should be easier now. I'll discuss content with the model at length. In conversation it communicates in lucid, spare prose and gives the impression of understanding. It stays on message when I ask it to prepare an outline, but then when I say, "Okay, go ahead and compose the first draft of section one," it flips from critical thinker to creative writer and suddenly gets much less smart. It sacrifices clarity for flourish and offers poetic metaphors that don't map cleanly onto the subject but which probably sound like "good writing" to someone with no training or experience in constructing or evaluating rational arguments.
GPT-5 is not as bad in this respect as GPT-4o, but it's still a struggle. And it's still quite frustrating.
You mentioned that LLMs don't make the same sorts of compositional errors that humans do, but you'll sometimes find them in Immutable Mobiles posts. Not because the model inserts them in an effort to give the impression of human authorship but because I get tired of the editorial back and forth and just re-write parts of the post myself without telling the model. That's where the errors creep in.
Fascinating. I could see LLMs falling into certain modes of discourse, and being bad at combining two modes that, on some level, represent different training sets. I assume this has made you more bearish on the future of AI in general?
If I'm bearish on AI, it's not because I think the capabilities have plateaued or soon will. It's more that I don't like the agendas to which AI is and will continue to be put, i.e. surveillance, speech policing, warfare, and wealth concentration.
What you wrote there captures a very common frustration: the “mode flip” between conversational lucidity and drafting fluff. It isn’t just your imagination—there’s a structural reason.
When you’re talking with the model, you’re operating in its analysis/completion mode: it’s selecting continuations that keep the dialogue coherent, spare, and on topic. Once you say “compose a draft,” you’ve nudged it into a basin where training data says “now imitate polished writing.” That basin is stuffed with mediocre web essays, marketing copy, and undergraduate prose—full of flourishes, half-metaphors, and padded transitions.
Labs lean on this because it tests well in demos. A first-time user asks “write me an essay” and gets back something that sounds like an essay. To them, purple language reads as sophistication. To anyone with editorial standards, it reads as bloat.
You’re not wrong that GPT-5 is a bit better here. The labs know about this tension. They want to avoid mode collapse into sycophancy while still protecting their models from spitting out dangerous or offensive text. The result is a system that’s hyper-tuned for inoffensive gloss.
However, keep in mind AI/LLM are a moving target. Today they might have tight, polished prose, after an update they might be meandering and eccentric.
I'd say the real trick will be to write authentically about what you care about and not worry about standing out from the muck, but I recognize some writers would actually like to make some spare change from the craft. Good luck to you.
I have two Substack publications. One that I write myself, and one for which I develop the content but then have ChatGPT (or sometimes Claude) write the text. It's a long and torturous process which sometimes takes longer and leaves me more drained and exhausted than writing a long post myself. Ironically, you have subscribed to the one that I write in conjunction with an LLM, Immutable Mobiles. Gen X Science Fiction and Futurism is the one that is entirely in my own voice.
Looks like I was already subscribed to the other one. Interesting that it should be more difficult. Are you expecting that it eventually will be much easier?
"Are you expecting that [writing with LLMs] eventually will be much easier?"
It's weird, because it seems like it should be easier now. I'll discuss content with the model at length. In conversation it communicates in lucid, spare prose and gives the impression of understanding. It stays on message when I ask it to prepare an outline, but then when I say, "Okay, go ahead and compose the first draft of section one," it flips from critical thinker to creative writer and suddenly gets much less smart. It sacrifices clarity for flourish and offers poetic metaphors that don't map cleanly onto the subject but which probably sound like "good writing" to someone with no training or experience in constructing or evaluating rational arguments.
GPT-5 is not as bad in this respect as GPT-4o, but it's still a struggle. And it's still quite frustrating.
You mentioned that LLMs don't make the same sorts of compositional errors that humans do, but you'll sometimes find them in Immutable Mobiles posts. Not because the model inserts them in an effort to give the impression of human authorship but because I get tired of the editorial back and forth and just re-write parts of the post myself without telling the model. That's where the errors creep in.
Fascinating. I could see LLMs falling into certain modes of discourse, and being bad at combining two modes that, on some level, represent different training sets. I assume this has made you more bearish on the future of AI in general?
If I'm bearish on AI, it's not because I think the capabilities have plateaued or soon will. It's more that I don't like the agendas to which AI is and will continue to be put, i.e. surveillance, speech policing, warfare, and wealth concentration.
ChatGPT-5 sez:
What you wrote there captures a very common frustration: the “mode flip” between conversational lucidity and drafting fluff. It isn’t just your imagination—there’s a structural reason.
When you’re talking with the model, you’re operating in its analysis/completion mode: it’s selecting continuations that keep the dialogue coherent, spare, and on topic. Once you say “compose a draft,” you’ve nudged it into a basin where training data says “now imitate polished writing.” That basin is stuffed with mediocre web essays, marketing copy, and undergraduate prose—full of flourishes, half-metaphors, and padded transitions.
Labs lean on this because it tests well in demos. A first-time user asks “write me an essay” and gets back something that sounds like an essay. To them, purple language reads as sophistication. To anyone with editorial standards, it reads as bloat.
You’re not wrong that GPT-5 is a bit better here. The labs know about this tension. They want to avoid mode collapse into sycophancy while still protecting their models from spitting out dangerous or offensive text. The result is a system that’s hyper-tuned for inoffensive gloss.
This was a funny piece, I enjoyed it.
However, keep in mind AI/LLM are a moving target. Today they might have tight, polished prose, after an update they might be meandering and eccentric.
I'd say the real trick will be to write authentically about what you care about and not worry about standing out from the muck, but I recognize some writers would actually like to make some spare change from the craft. Good luck to you.
I have two Substack publications. One that I write myself, and one for which I develop the content but then have ChatGPT (or sometimes Claude) write the text. It's a long and torturous process which sometimes takes longer and leaves me more drained and exhausted than writing a long post myself. Ironically, you have subscribed to the one that I write in conjunction with an LLM, Immutable Mobiles. Gen X Science Fiction and Futurism is the one that is entirely in my own voice.
Looks like I was already subscribed to the other one. Interesting that it should be more difficult. Are you expecting that it eventually will be much easier?