Editing AI's Japanese draft barely beats writing it myself
This is an honest episode. Up to now I’ve been writing enthusiastically about “turning AI into a team and making a full-length novel,” but once I entered the Japanese revision and editing phase, I came back down to earth.
To cut to the conclusion, the feeling was: “Isn’t this, in the end, just slightly easier than writing it myself?”
For anyone skeptical of AI writing, I’ll be honest: this work was extremely hard, hard enough that I found myself thinking writing the whole thing myself might have been easier. …Well, that’s not actually true. Ha. But it was grinding, laborious work, enough that the thought crossed my mind.
What “just slightly easier” actually meant
AI gives you a first draft at tremendous speed. That part really is fast. But whether the Japanese that comes out is good enough to publish as is is an entirely different question.
- Unnatural phrasing
- Quirky punctuation habits
- Dropped particles and sentence endings
- A distinctly “AI-sounding” style creeping in at odd moments
For example, a sentence like this slips in without a second thought.
Stephen, outside the circle of players, holding his notebook, stood there.
Too many commas, and the sentence keeps catching its breath in fragments. I’d tighten this to something like:
Stephen stood outside the circle of players, holding his notebook.
Other issues turned up too: paragraphs that should have flowed were chopped into short clipped sentences, ”***” scene-break lines overused even when there was no real scene change, match results dropped in as bare nouns like “3 to 1.” Each one was small on its own, but left unaddressed, they would steadily drag the whole thing down in quality.
Checking and fixing these one sentence at a time ends up being entirely my own job. I can’t hand this judgment off to AI. In the end, a human decides whether a given piece of Japanese is acceptable or not.
In other words, the labor of writing from scratch goes down, but the labor of guaranteeing quality barely decreases at all. If anything, a different kind of exhaustion gets layered on top: reading through and fixing text someone else wrote. That turned out to be the real substance behind “just slightly easier.”
In the end, QA (re-reading and fixing) was the heaviest part
Having published a book myself, I already knew this, but writing as a craft takes the most effort in QA (re-reading and fixing). One novelist once said, “Writing takes a week, fixing it takes two or three months,” and this time I felt that in my bones.
In fact, the survival rate of the text first generated from the synopsis and plot had already dropped below 70% by the middle of editing. The original shape kept getting worn away and rewritten. Even if AI can mass-produce first drafts, nobody escapes this “fixing” process — that’s the most honest thing I took away from this book.
Make it write at length, and the style slowly erodes
Another headache was that during a long generation session, the style gradually drifts. The opening chapters are clean, but the further in you go, the more subtly it breaks down. In fact, I picked the opening chapters, where the style was most stable, as my “standard,” and put a dozen-plus later chapters through a full rewrite to bring them in line with the style of the opening chapters. Partial fixes couldn’t keep up. Partway through, I gave the AI I was working with this instruction:
Only checking the opening and the end of a chapter is risky, so go through the entire text properly.
I wrote that line to discipline my own urge to take the easy route and only check the key spots. In the end, “just check the beginning and the ending” turned out to be a complete trap: there was no way around reading every single sentence of the whole text.
I turned things into strict rules too. Take something as small as a character’s tone of voice: “Don’t drop particles in narration,” “James’s dialogue can drop particles, but don’t let noun-ending sentences run in a row,” “Stephen’s particles should be precise except when he’s excited,” “Match results should be in flowing sentences, not noun-only endings,” “Divider lines only at a meaningful transition” — I spelled out judgment criteria like this in words and handed them to the AI. As I wrote last time, the AI won’t deliver unless you codify things to this degree.
What I did to keep the editing moving anyway
Even though I’d cooled off, I couldn’t just abandon it, so I restructured how I worked.
- Separate “detection” from “correction” (first have it flag problem spots, 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 each rule to a concrete before/after example
- For each chapter, create a review Markdown, and judge each item with checkboxes: “adopt / leave as is / fix myself”
Once I switched to this approach, not expecting a single masterstroke and instead breaking the work down into a mechanical process, things finally stabilized. Put another way, it was unusable unless I designed the process to that degree.
Reading this far, it probably comes across as 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.” Same AI, but it hit the mark in a completely different place.
Originally published in Japanese at https://clazytech.com/2026/06/1645/. Translated with LLM assistance and reviewed before publication.