AI Cuts Only a Little Work From Japanese Editing, Not the Hard Part
This is an honest installment.
Up to now I’ve written with plenty of swagger about “making AI into a team and producing a full-length novel,” but once I entered the Japanese revision and editing phase, I cooled off for real.
To cut to the conclusion, the feeling was, “Isn’t this, in the end, only slightly easier than just writing it myself?”
For anyone with reservations about AI writing, let me be honest: this work was extremely hard, hard enough that at times I wondered whether writing it all myself would have been easier. That’s obviously not true. Ha. But it was steady, grinding work, hard enough that the thought crossed my mind.
What exactly was “only slightly easier”
AI gives you a first draft at tremendous speed. That part really is fast. But whether the resulting Japanese was quality you could put out into the world as-is was an entirely different question.
- unnatural phrasing
- idiosyncratic habits in punctuation
- dropped particles and word endings
- the “unmistakably AI” style that creeps in at odd moments
For example, a sentence like this would slip in without a hint of self-awareness.
Stephen, outside the circle of players, holding his notebook, stood there.
Too many commas, and the sentence keeps stopping for breath in fragments. I’d tidy it up to something like
Stephen stood outside the circle of players, holding his notebook.
Other problems came up too: paragraphs that should have flowed were chopped into short fragments, ”***” section breaks overused even where there was no real scene change, match results dropped in as a bare noun like “3–1.” Each one was small on its own, but left alone, they would steadily drag down the quality of the whole thing — that was the pattern.
Going through and fixing this line by line ends up remaining entirely my own job. I can’t hand this judgment off to AI wholesale, because in the end, a human decides whether a given piece of Japanese is acceptable or not.
In other words, the labor of writing from zero goes down, but the labor of guaranteeing quality barely goes down at all. If anything, a different kind of exhaustion piles on top: reading through everything someone else wrote and fixing it. That was the truth behind “only slightly easier.”
In the end, QA (re-reading and correcting) was the heaviest part
Since I’ve published a book myself, I understand this: with writing, QA (re-reading and correcting) is by far the most labor-intensive part. A certain novelist once said, “Writing takes a week; fixing it takes two or three months,” and this time I felt that truth in my bones.
In fact, of the text first generated from the synopsis 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 took from this book.
Let it generate at length, and the style slowly erodes
Another troublesome thing: over a long generation session, the style drifts little by little. The opening chapters were clean, but the further in you went, the more subtly it fell apart. I ended up designating the earliest chapters, where the style was most stable, as the “standard,” and put the latter dozen-plus chapters through a full rewrite to bring them in line with the style of the early chapters. Partial touch-ups couldn’t keep pace. Partway through, I gave the AI I was working with an instruction like this:
The idea of only checking the opening and the end of each chapter is dangerous, so please go through and check the full text properly.
It’s a line I wrote to check my own temptation to cut corners and only look at the key spots. In the end, “just check the opening and the close” turned out to be a complete trap. Reading every single sentence of the full text was the only way.
I also enforced this through explicit rules. Take a single character’s speech patterns, for instance: “never omit particles in narration,” “James’s lines can drop particles, but not consecutive noun-ending sentences,” “Stephen’s particles should be exact except when he’s excited,” “match results should be in flowing sentences, not just a bare noun,” “section breaks only at meaningful transitions,” and so on. I put the judgment criteria into words and handed them to the AI. As I wrote last time, unless you codify things to this degree, AI won’t respond the way you need.
What I did to keep the editing going anyway
Even having cooled off, I couldn’t just abandon it, so I restructured the approach.
- Separate “detection” from “correction” (first have it flag problem spots, fix them in a separate pass)
- Restrict output to diff format (don’t have it regenerate the whole text)
- One rule per pass (don’t have it fix multiple things at once)
- Always tie each rule to a concrete Before/After example
- Build a review Markdown file per chapter, with checkboxes for each item to judge “adopt / leave as is / fix myself”
Once I switched to this approach, not expecting some all-in-one feat of virtuosity but breaking it down into a mechanical process, things finally stabilized. Put the other way around, this also means it was unusable unless I designed the process to this extent.
Reading this far, it probably looks fairly harsh. But next time is the swing back. I’ll write about how “in the KDP release work, generative AI genuinely helped me in a big way after all.” Even with the same AI, where it hit home was completely different.
Originally published in Japanese at https://clazytech.com/2026/06/1645/. Translated with LLM assistance and reviewed before publication.