Yep, it's like Coke Zero vs Diet Coke: 10x the flavor and 10x the calories.

You can say anything to an LLM, but it’s not going to actually write in your voice. When I was writing a very long blog post about “creative writing” from AIs, I researched Sudowrite briefly, which purports to be able to do exactly this; not only could it not write convincingly in my voice (and the novel I gave it has a pretty strong narrative voice), following Sudowrite’s own tutorial in which they have you get their app to write a few paragraphs in Dan Brown’s voice demonstrated it could not convincingly do that.

I don’t think having a ML-backed proofreading system is an intrinsically bad idea; the oft-maligned “Apple Intelligence” suite has a proofreading function which is actually pretty good (although it has a UI so abysmal it’s virtually useless in most circumstances). But unless you truly, deeply believe your own writing isn’t as good as a precocious eighth-grader trying to impress their teacher with a book report, don’t ask an LLM to rewrite your stuff.

No man. This is the whole problem. Don't sell yourself short like that.

What is a writing "voice"? It's more than just patterns and methods of phrasing. ChatGPT would say "rhythm and diction and tone" and word choice. But that's just the paint. A voice is the expression of your conscious experience trying to convey an idea in a way that reflects your experience. If it were just those semi-concrete elements, we would have unlimited Dickens; the concept could translate to music, we could have unlimited Mozart. Instead—and I hope you agree—we have crude approximations of all these things.

Writing, even technical writing, is an art. Art comes from experience. Silicon can not experience. And experiencers (ie, people with consciousness) can detect soullessness. To think otherwise is to be tricked; listen to anything on suno, for example. It's amazing at first, and then you see through the trick. You start to hear it the way most people now perceive generated images as too "shiny". Have you ever generated an image and felt a feeling other than "neat"?

Only if you have a very low bar for what constitutes "in your voice".

Just ask it to write "in the style of" a few famous writers with a recognizable style. It just can't do it. It'll do an awfully cringe attempt at it.

And that's just how bad LLMs are at it. There's a more general problem. If you've ever read a posthumous continuation of a literary series by a different but skilled author, you know what I mean.

For example, "And another thing..." by Eoin Colfer is written to be the final sequel to the Hitchhiker's Guide, after Douglas Adams died. And to their absolute credit, the author Eoin Colfer, in my opinion, pretty much nails Douglas Adams's tone to the extent it is humanly possible to do so. But no matter how close he got, there's a paradox here. Colfer can only replicate Adams's style. But only Adams could add a new element, and it would still be his style. While if Colfer had done exactly the same, he'd have been considered "off".

Anyway, if a human writer can't pull it off, I doubt an LLM can do it.

I have tried this. It doesnt work. Why? A human’s unique style when executed has a pattern but in each work there are “experiments” that deviate from the pattern. These deviations are how we evolve stylistically. AI cannot emulate this, it only picks up on a tiny bit of the pattern so while it may repeat a few beats of the song, it falls far short of the whole.

This is why heavily assisted ai writing is still slop. That fundamental learning that is baked in is gone. It is the same reason why corporate speak is so hated. It is basically intentional slop.

Best case scenario, this means writing new blog posts in your old voice, as reconstructed by AI; some might argue this gives your voice less opportunity to grow or evolve.

I think no, categorically. The computer can detect your typos and accidents. But if you made a decision to word something a certain way, that _is_ your voice. If a second party overrides this decision, it's now deviating from your voice. The LLM therefore can either deviate from your voice, or do nothing.

That's no crime, so far. It's very normal to have writers and editors.

But it's highly abnormal for everyone to have the _same_ editor, famous for the writing exactly the text that everybody hates.

It's like inviting Uwe Boll to edit your film.

If there's a good reason to send outgoing slop, OK. But if your audience is more verbally adept, and more familiar with its style, you do risk making yourself look bad.

> especially if it's an LLM with extensive memory of your writing.

Personally I'm not submitting enough stuff to an LLM to give it enough to go on.

Exactly. It's so wild to me when people hate on generated text because it sounds like something they don't like, when they could easily tell it to set the tone to any other tone that has ever appeared in text.