AI Writing Only Shifts The Labor, It Doesn't Remove It
This is an honest installment. Up until last time I wrote confidently about “turning AI into a team to produce a full-length work,” but once I entered the Japanese proofreading and editing phase, I sobered up for real.
To cut to the conclusion — the feeling was, “Isn’t this, in the end, just slightly easier than writing it myself?”
For anyone with a bias against AI-assisted writing, I’ll be honest: this work was extremely hard, hard enough that I sometimes thought it might have been easier to write the whole thing myself. …Well, that’s not quite true, of course. Ha. But it was steady, grinding work, hard enough that I kept wanting to blurt that out.
What exactly was “just slightly easier”
AI produces a first draft at tremendous speed. That part really is fast. But whether the Japanese that comes out is quality you can publish as-is is an entirely different matter.
- Unnatural phrasing
- Idiosyncratic habits in comma placement
- Dropped particles and sentence endings
- The occasional stretch that’s unmistakably “AI-flavored” prose
For example, a sentence like this would slip in without a hint of self-awareness.
Stephen stood, outside the circle of players, holding his notebook.
There are too many commas, and the sentence keeps catching its breath in fragments. I’d clean it up to something like:
Stephen stood outside the circle of players, holding his notebook.
There were other issues too: paragraphs that should have flowed were chopped into short, disconnected sentences; the ”***” section break overused even where there was no real scene change; match results dropped in as a bare noun phrase like “3-1.” Each issue on its own was small, but left alone, they would reliably drag down the overall quality.
The work of checking and fixing these sentence by sentence remains entirely my job. I can’t hand this judgment off to AI wholesale, because in the end, it’s a human who decides whether a given piece of Japanese is acceptable.
In other words, the labor of writing from zero goes down, but the labor of guaranteeing quality barely decreases at all. If anything, a different kind of exhaustion piles on top — the work of “reading everything someone else wrote and fixing it.” That was the true nature of “just slightly easier.”
In the end, QA (rereading and revising) was the heaviest part
Having published a book myself, I know this: with writing, QA (rereading and revising) is by far the most labor-intensive part. A certain novelist once said, “Writing takes a week; fixing it takes two to three months.” This time I felt that in my bones.
In fact, of the text initially generated from the outline and plot, the survival rate had already dropped below 70% by the midpoint of editing. The original form kept getting whittled away and rewritten. Even if AI can mass-produce first drafts, nobody escapes this “fixing” process — that’s the most honest realization I got from this book.
Write long, and the style slowly erodes
Another troublesome thing: if you keep generating over a long session, the writing style gradually drifts. The early chapters are clean, but the further in you go, the more subtly it degrades. In fact, I picked the early chapter with the most stable style as my “baseline,” and put a dozen-plus later chapters through a full rewrite to align them with that early chapter’s style. Partial touch-ups couldn’t keep up. Partway through, I gave the AI working on this an instruction like the following:
The idea of only checking the opening and the chapter ending is dangerous. Please go through and check the entire text properly.
That’s a line I wrote to discipline myself against the temptation to cut corners and check only the key spots. In the end, “you only need to look at the opening and the close” turned out to be a complete trap. There was no way around reading the entire text, not skipping a single sentence.
I also enforced strict rules. Take a character’s speech patterns alone: “don’t omit particles in narration,” “James’s dialogue can drop particles, but don’t string together consecutive noun-ending sentences,” “Stephen should use particles accurately except when excited,” “match results should be written as flowing sentences, not noun-endings,” “section breaks only for meaningful transitions” — rules like these. As I wrote last time, unless you codify things to this level of detail, AI won’t respond properly.
Ways I found to keep the editing moving anyway
Even though I’d sobered up, I couldn’t just give up, so I restructured the approach.
- Separate “detection” from “correction” (first have it flag problem spots, then fix them in a separate step)
- Restrict output to diff format (don’t have it regenerate the whole text)
- Run one rule per pass (don’t have it fix multiple things at once)
- Always tie a rule to a concrete Before/After example
- Create a review Markdown per chapter, with checkboxes for “adopt / leave as-is / fix myself” on each item
Only after switching to this approach — don’t expect a single masterstroke to do it all; break it down into a mechanical process — did things finally stabilize. Put another way, it was unusable unless I engineered the process to this degree.
Reading this far, it might look fairly harsh. But — next time is the swing back the other way. I’ll write about how “in the KDP release work, generative AI genuinely helped me a lot after all.” Same AI, but it hit home in a completely different place.
Originally published in Japanese at https://clazytech.com/2026/06/1645/. Translated with LLM assistance and reviewed before publication.