Starting around high school, people began asking me to help proofread their writings. Occasionally it’d be a college essay or a publication, but I also helped with simple things like emails or blog posts or real estate listings. Eventually, by the time I transitioned into engineering management, a good part of my job consisted of reading and commenting on RFCs (Request for Comment docs) and technical specs.
I learned to ask what was expected by my readership. Was it to provide high-level critique and clarifying questions? Should I look at the logic of the proposal and whether I agreed with its premise? Do I correct grammar and spelling and diction?
A handful of times, I was asked to rewrite a very rough draft into something that “didn’t suck.” I was too nice and acquiesced to their request, but my corrections got more subtle as my editing improved. Once, as I was reviewing my edits with a friend, he asked me why I left his awkward sentences as they were, even as I was rephrasing them in my own words to confirm his intentions. I told him that my job was to make the writing better—in his voice, not mine.
The book On Writing Well touches on this aspect of the written word. Structure and grammar make for readable prose, but those facets alone do not develop character. How someone speaks—and writes—is a reflection of their personality, their experience, and their unique perspective on the world.
To give a simple example, I read papers from Chinese-speaking friends and family, for whom English is a distant second language in fluency. The difference in sentence composition, the lack of pluralities and the confusion between past and present tense, give away their internalized translation patterns. It’s in the same vein as badly-translated Chinese restaurant menus, though of course nowhere near as egregious.
So I’ll usually correct these simple mistakes, but sometimes I make it a point to leave a sentence alone, even when it’s technically incorrect. Most people read by sounding the words in their head, a method called subvocalization, that helps understanding and retention1. As long as the point comes across, retaining that voice is more important than by-the-book grammar, if small imperfections help them hear the author.
This aspect of writing character is what’s missing in LLM-generated prose. LLM text, by its nature as averaged model, will always sound correct—exactly what we were taught as students of English over 12+ years of primary and secondary schooling. The paragraphs have an exacting structure; they present logically and pleasantly, with extraneous words stripped out, and sentences evenly sized.
In revamping my own writing process, I’m now interacting with ChatGPT more to ideate and iterate on ideas. I ask the AI to generate rough drafts that wrap nascent thoughts into fully-formed paragraphs, and it dutifully spits out posts that are honestly pretty easy to follow. It generates heaps of material to react to, easily enumerates possibilities from multitudes of positions, and speeds up research immensely. It’s a blogging coach, a rendition of the idealized AI tutor that Salman Khan described in Brave New Words for students.
But the generated text is bland enough to be forgettable, and I end up rewriting all of it by the time it gets published. I like the quirks and offhand references in these posts, and their inclusion is also a major part of my writing which has yet to be captured by LLMs. ChatGPT can iterate to a final draft by itself, and emulate some surface-level characteristics of an author’s tone2. If the goal is to present an argument in a straightforward fashion, then this is a reasonable workflow that also optimizes for clarity and time.
It has yet to figure out how to mimic personality, experience, and perspective.
A strategy to read faster is to speed up or even eliminate this habit of subvocalization, though its efficacy is controversial.↩
I have tried pointing LLMs directly to this blog, just so it’d have enough data for its context window.↩
