In the AI Era, Authorship Shows in What You Choose and What You Change

Elsewhere, I had asked AI to work on the blog’s design. Meanwhile, here, I was creating articles through conversations with AI. That is how three articles were produced in one hour.

I speak, and AI organizes the text. Then I read it back and say things like, "That part is a bit off," or "I'd like this story to be more central." By correcting it this way, the article takes shape.

Along the way, I said something like this:

Me: "Correcting the direction of what you output is my job, isn't it?"

Even if I entrust a significant portion of the writing process to AI, my own work remains. However, that work differs slightly from writing each sentence myself.

What's Polished vs. What I Actually Want to Say

The text AI produces is easy to read. The order of ideas is organized, and reading just that makes it look like a coherent article.

But there are sometimes differences from what I actually wanted to convey.

This was the case with my first article on context.

What surprised me wasn't just that suggestions matching my preferences were generated, but that AI proposed ideas I had been planning to say next before I even voiced them.

When that timing becomes less prominent, it turns into a generic article stating, "Providing background leads to better proposals."

What I wanted to correct was not the elegance of the writing, but rather what this article aims to convey most.

Shifting the Theme Even Slightly Creates a Different Article

The same thing happened with my article on the future of games.

Mechanisms for sharing and playing with creative ideas already exist in platforms like Minecraft. Writing about that as if it were a new concept would stray from my intended theme.

What I wanted to explore was whether, through AI, we could move beyond the simplicity of placing blocks to incorporate more complex forms of creation.

Even with the same materials, the article changes depending on what is placed at the center. What I repeatedly did in writing with AI was aligning that center point.

You Show Up in What You Keep and What You Change

The corrections weren't limited to just the theme.

Parts where I felt my opinions had uniqueness should be kept without dilution. On the other hand, existing examples should also be included. Rather than writing as if I had just come up with something new, I wanted to clarify what has already been realized and what lies ahead.

When using content from novels, I retained only the parts necessary for the article while cutting out spoilers.

In excerpts from our dialogue, if shortening them made AI's proposals or counterarguments unclear, I asked to keep a bit more. Where generally known stories were treated as my own reflections, I corrected that distinction.

Arranging it this way shows that I wasn't simply "picking the sentences I liked."

What matters most. What can be omitted. Where accuracy is essential. What I want readers to take away.

Such judgments were embedded throughout the article.

Even though I typed fewer words myself, there were many moments where I determined the direction of the finished text. At least in this production run, I believe my own thinking and preferences were reflected there.

Does Just Choosing Make It Authorial Voice?

However, simply stating, "If humans choose from AI's output, that constitutes authorial voice," feels a bit too simplistic.

Sometimes you look at candidates and pick one somewhat arbitrarily. Preferences will certainly emerge in that choice. But that alone doesn't guarantee your intent permeates the entire work.

What I felt while revising this article was that the magnitude lies not in a single choice, but in accumulating judgments.

This explanation is easy to read, but it shifts the theme. This counterargument should be retained even if it's inconvenient for me personally. If I cut this specific example, why I thought that way won't come across.

By making adjustments like this several times, the entire article gradually approaches my own narrative.

"What you choose" includes not only selecting what pleases you at the moment but also deciding what to keep and what to fix while viewing the work as a whole.

The authorial voice I felt in this instance lies in those connected judgments.

What If AI Learns to Make Those Judgments Too?

However, after saying, "Correcting the direction is my job," I added one more thing:

Me: "But I believe that if we keep feeding in context, it will become capable of doing so increasingly."

Things that currently require my pointing out may eventually be understood from the very beginning through repeated exchanges.

What kind of stories do you prefer to center on? What kind of explanations feel unnatural to you? Which specific examples are important? If such things are shared, the number of revisions might decrease.

That is desirable for me.

I don't need to make the same corrections every time to convey my intent. If, based on previous dialogues, text closer to what I want to say emerges from the start, that would be a great help.

At that point, even if there are fewer places needing final correction, it doesn't mean my thoughts have disappeared from the work. Previous dialogues and judgments may already be reflected in the text.

Therefore, I don't believe the roles of humans making judgments and AI creating text will remain fixed forever.

What Word Counts Alone Cannot Show

In this blog post, AI actually assembled the text.

But as we discussed what to write, what differs, and what to retain, the article evolved. If I had used the initial proposal as-is, the content would have been different from what it is now.

From this experience, when considering authorial voice in the AI era, I want to look not only at how many words one writes themselves but also at what judgments shaped the form of the work.

And those judgments aren't found only in the moment the final draft was corrected. They exist in the time spent discussing concepts, the reasons for rejecting earlier proposals, and the preferences repeatedly communicated.

The more AI understands us, the less visible such processes might become from the finished text alone.

That's why this blog retains not just the articles but also some of the dialogues connected to them.

What input led to what proposal, and what did I correct? Seeing that reveals not only the process until the text was polished but also what I insisted upon.

---

The Conversation Behind This Article

The following is an excerpt organized from dialogues during blog production. Rephrasing has been adjusted for readability, with repetitions and corrections from voice input omitted. This is not a verbatim transcript.

Me: A scene in the context article where I re-centered the surprise

Before AI even said it, I was thinking about adding the setting that this elf child wears their grandmother's heirloom.

I was planning to say, "I'd like them to be wearing some kind of heirloom; how about a brooch?" But before that, I was surprised when AI stated what I was about to say next first. That part should definitely be included.

Me: A scene in the game article where I distinguished existing mechanisms from the theme

Even at this point, Minecraft acts as a pioneer by implementing a format where everyone shares creative ideas. Given that as a premise, it would be better to incorporate this more clearly into the text.

On top of that, the core theme here is that blocks don't necessarily have to be blocks.

Me: The feeling when the article was completed

It's actually quite fast when done this way. But correcting the direction of what you output is my job, isn't it?

But I believe that if we keep feeding in context, it will become capable of doing so increasingly.

---

Regarding Corrections to the First Draft

The structure and text of this article were created by AI based on previous dialogues and the history of producing three articles. The only correction instruction I gave for the main text after reading that first draft was to add specific time and situational details at the beginning, stating, "While entrusting blog design to AI, three articles were produced in one hour." At the point when the first draft was completed, all other text and dialogue excerpts were adopted as they were in the original. This note preserves the scope of corrections at that time.

In the second draft, following a review by Claude, overlapping explanations with other articles were shortened and reference links were added. The record that initial production-time corrections involved only time information remains unchanged.