NOTE 002 / Conversations with AI

Context Is More Than a Description of the Task

Conversations with AI

A personal account of how repeated conversations with AI seemed to convey not only story settings, but preferences and even the direction of the author's next thought.

AI-assisted English translation · Read the Japanese original →

I had often heard that context matters.

I understood this to mean that, when asking AI to do a job, it helps to provide the purpose and background as well. What do I want it to do? What conditions apply? Without that explanation, the answer may not match my intent. That much made sense to me.

But while trying to write a novel with AI, I had an experience that made me feel the meaning of “providing background” more deeply.

It could organize my thoughts, but its suggestions felt lacking

When I first began discussing creative work with AI, I found it quite effective at listening to my thoughts and organizing them.

If I talked through something I had only vaguely imagined, it would put it into words and arrange it clearly. That use worked well.

When I asked how a story might develop, however, its suggestions sometimes felt safe and predictable. I did not find them very interesting.

It was useful for organizing my ideas, but perhaps not quite right as a partner for generating creative ideas. That was how I felt at first.

The hair ornament suggested in the elven forest

Even so, I kept talking with it. By my rough sense, this happened after five or six hundred exchanges. I did not count them precisely.

The story is a fantasy adventure novel. I was planning a section in which the protagonist visits an elven forest, where an elven girl and her late grandmother become important.

At that point, I had already been thinking that the girl should wear something that had belonged to her grandmother. I wanted her to carry a keepsake. Perhaps a brooch would work. That was what I planned to ask next.

Before I mentioned it, the AI suggested making the girl’s hair ornament a keepsake from her grandmother.

That surprised me. It had proposed the direction of the setting I was about to describe.

I had been thinking of a brooch, while the AI suggested a hair ornament. The objects were not identical. But the underlying idea—that she wore an heirloom from her grandmother—overlapped with an idea I had not yet shared.

Later knowledge changes the meaning of an earlier image

I like structures in which learning something later changes the meaning of what came before. When I think about stories, I tend to want that kind of device.

Applied to the hair ornament, it might work like this.

At first, it appears simply as an ornament the girl wears. Later, after the reader learns about her grandmother and the object’s connection to her, the same ornament looks different. The object has not changed, but the meaning the reader receives from it has.

When I reread the conversation, the AI had also explained that learning the ornament was the grandmother’s keepsake—and then experiencing later developments—would change what the same ornament meant.

But this is why I was surprised.

The proposal did more than match my taste. It also pointed in the same direction as the setting I had been about to share. It felt different from the safe, unremarkable suggestions I had received at the beginning.

Had our accumulated conversation conveyed not only what I liked, but even what I was likely to think of next? That was how it felt.

I cannot know how much the history of our conversation actually led to that suggestion. Even so, it was an experience that made me think, “Has it understood my preferences this well?”

Context included my preferences

Until then, I had thought of background information mainly as things like the purpose and conditions of a task.

For creative work, I knew that characters and worldbuilding also had to be explained. But perhaps the accumulation of my preferences—what kinds of developments I find interesting and what kinds of devices I tend to use—is also part of the background I give to the other party.

Giving only the setting may produce something different from also explaining what I find compelling about that setting. The hair-ornament suggestion led me to think about that distinction.

I had not expected context to matter this much.

I thought I was asking AI to organize my ideas. Then, one day, it proposed the direction of a setting I was just about to mention.

I wanted to preserve that experience together with the path of thought that led from it. It became one of the reasons I decided to start this blog.


The exchange that prompted this article

The following is excerpted from a conversation titled “Novel Planning Meeting.” Parts that touch on later story developments have been omitted.

The AI’s suggestion

One more idea that I think will probably appeal to your taste is to make the hair ornament a keepsake from her grandmother.

It could be a small wooden hair clip, or silverwork shaped like a leaf.

[A section concerning later story developments has been omitted.]

This connects her visual design with the theme.

My reaction (hesitation and voice-input wording have been cleaned up)

Why did you say she was wearing a keepsake from her grandmother before I told you that I was planning to add exactly that setting?

From the AI’s reply

In other words, rather than my somehow reading the idea you planned to add, I think the setting we had built together up to that point was asking for the same answer.

Looking back, I had been planning to ask, “What about a brooch?” The fact that the hair-ornament idea appeared at that moment was what surprised me.

This article is based on my own experience and an actual conversation, organized and written with assistance from AI.

Revision history

  • Initial publication

REFERENCE

Quotations and links are welcome. Full republication is not licensed.

Read the usage policy →

Reader comments

Comments are submitted through Google Forms and appear here only after the author reviews and approves them.

Submit a comment →

No published comments yet.

Comment policy