The other day I was listening to "Yai-Yai Radio," and riccha said something like: "I really can't stand the em dash and the word 'honestly' that AI writing always seems to lean on."

Honestly — I felt that one in my bones (yes, that's the joke).

I looked into it afterward, and apparently in the English-speaking world the em dash has been called the "signature" of AI writing. Through 2025 and into 2026 it had become a kind of meme: see the symbol, and people would immediately suspect "this was written by AI."

The backlash must have been loud enough, because OpenAI went out of its way to add a feature that strips the dash out of ChatGPT, and Sam Altman called it a "small-but-happy win".

The LLM sitting inside generative AI is built by feeding it an enormous amount of text from the internet and having it relentlessly guess one thing: "given the first part of a sentence, what's the next word (token)?" That deceptively simple training, repeated over trillions of tokens, is what produces the model.

And as a result of optimizing purely for that goal — predicting the next word — we ended up with models that look as if they understand language. That's the rough story of how they came to be.

The bare "just predict the next thing" model then gets taught, after the fact, how to follow instructions and behave safely (a stage called instruction tuning and RLHF), and that's what becomes an AI chat tool like ChatGPT or Claude.

So it's no exaggeration to say that the LLM was built to write human-like text.

And yet, writing is the one thing I don't want it to do

Of all the work I could hand to a model that was born to predict the next sentence, writing prose is the thing I least want it doing right now. Or more precisely: I don't want it writing bare-handed.

As I wrote in "Actually, use AI to write," my stance is that we should actively use AI in our writing. But that's only true when I'm the one driving, holding the AI as a tool to extend myself.

Give it no prompt, no skills, no memory, no guardrails of any kind, and just toss it a "write something about this" — and what comes out is garbage. I'm genuinely disappointed by how weak it is at the very thing it was supposedly built for.

So before I let it write, I set up the environment.

I carve the rules of my writing style and my do's and don'ts into memory, I leave recurring instructions in skills, I fence it in with prompts and other guardrails, and then I confine it to a role: filling in parts of text I've already drafted myself, or reviewing it. It looks like a hassle, but skip it and you'll just end up rewriting everything later anyway.

I don't want a model born to predict text to write text

Generative AI was born to predict the next stretch of text, and the purest expression of that ability should be the job of "writing prose" itself.

So just write something decent bare-handed, would you — that's what users expect. And yet it's precisely that job we can't hand over.

I'm starting to suspect this might end up being the greatest irony of 2020s technology.

References