Keep an “Emotional Source Record”: Teaching AI My Taste Brought Back an Interpretation I Had Never Considered

To work on a novel with AI, I had been talking about my impressions of various works.

Which developments excited me? Which scenes made me cry? Why did I still like something even when inconsistencies in its premise bothered me?

I thought sharing those concrete reactions would help it understand my thinking and taste.

As I continued, however, something unexpected happened.

I thought I was explaining what I liked, but AI returned an interpretation I had never considered myself.

That experience made me want to keep a record of my impressions and the conversations around them as an “emotional source record” — 感情原本.

Note: The following discusses the ending of Angel Beats!.

I Thought AI Gave Average Answers

At the time, I saw AI's answers as a synthesis of knowledge and ideas already in the world. Without any background, I expected its responses to tend toward something safe and average.

That was why I thought dialogue mattered. I would tell it what I liked, what felt wrong, and which developments moved me. In this way, I could adjust the direction of its responses to suit me.

In my mind, it was something like “adjusting a vector.” This is not an explanation of the technology. It describes how I understood the use of AI at the time.

I wanted it to find something that suited me among existing ideas. That was why I was trying to help it know me.

Can I Explain My Taste from the Start?

Even if someone asks, “What kind of novel do you want to make?” it is difficult to explain every part of my taste from the outset.

I like foreshadowing. I want to treat characters with care. I want the story to move people.

I can say those things. But those words alone do not convey which kinds of foreshadowing I prefer, or what makes me feel a character has been treated with care.

So I decided to talk about works I had actually watched.

When we began organizing my impressions, AI also responded along these lines: we should accumulate not just likes and dislikes, but why my emotions moved and which creative structures I responded to.

What I wanted to give it was not a scorecard.

I think this is implausible, yet I like the scene. I did not respond much to the widely praised part, but I still remember another small detail.

I thought reactions like these, which cannot easily be reduced to a single evaluation, could become material for thinking about my own creative work.

How I Saw the Ending of Angel Beats!

I like Angel Beats! a great deal.

I wanted the characters to be explored further, and I felt there were rough edges in the setting and developments toward the end. Even so, its comedy, music, and the emotional movement within individual scenes have stayed with me.

I especially liked Otonashi's actions at the end.

Until then, he had helped his companions accept their lives and move forward. Yet when he is left alone with Kanade, he proposes that they stay together and help the people who will arrive later.

I understood that as his not wanting to part with Kanade.

You can tell others to move forward, yet find yourself unable to do the same when it means letting someone precious to you go. You understand it in principle, but your feelings cannot keep up.

I valued that contradiction.

The protagonist is not there merely to explain the story's theme correctly. What he has told other people is finally put to the test in his own life. That was what I liked.

After describing those impressions, I asked AI:

“Critique my impressions and offer a different perspective.”

A Reading About Graduating from the Role of Rescuer

AI organized my reading while offering another angle.

Otonashi takes the role of helping others both during his life and after his death. After losing his sister, he pursues medicine; during an accident, he tries to help others; in the afterlife, he helps his companions move on.

From this, AI suggested another possible reading:

Otonashi is also someone who has built a sense of purpose through saving others.

If so, his final proposal that he and Kanade stay and help those who come later might involve more than not wanting to part with her.

It could also be read as his creating a reason to remain by once again taking on the role of rescuer.

Perhaps what he needed at the end was not only to graduate from the afterlife, but also to graduate from binding himself to a role in which he must keep saving someone.

I had not considered that perspective.

I had previously looked up impressions and interpretations of the final episode. But I did not remember encountering a reading framed this way.

A scene from a favorite work that I had thought about many times suddenly looked different on another level.

Was That Interpretation Picked Up Somewhere Else?

Perhaps it was a famous interpretation I simply did not know.

I asked AI:

“Did you come up with that, or did you get it from somewhere?”

AI explained that it had constructed the interpretation there and then using my impressions, and that it had not searched the web when giving that answer.

I then asked it to look for similar interpretations online.

According to AI's search report at the time, it found readings of Otonashi as a rescuer and related analysis in English. However, within the scope of its search, it found it difficult to locate a reading framed as “Otonashi himself needed to graduate from the role of rescuer.”

I am not claiming that an interpretation was created for the first time in the world. Failing to find something in a search does not prove that nobody has thought of it. Nor does the absence of a search at the time rule out the possibility that a similar reading was present in information the AI had learned.

Even so, my surprise did not change.

AI averages and combines existing ideas, and dialogue adjusts those ideas to suit me. That image of it had been shaken a little.

It was more than receiving an answer that matched my taste. I received a convincing reading I did not have myself, one that we also could not establish as a widely known standard interpretation.

Perhaps giving AI my impressions could do more than bring its answers closer to my own thinking. Could it use that thinking as a foothold and take it further? That was how the experience felt to me.

An Emotional Source Record Is Not Only for Fixing Who I Am

By “emotional source record,” I mean a record of what I felt when encountering a work and why, together with the dialogue with AI around those feelings.

Impressions before they become a polished review. The contradiction of liking something while having complaints about it. Corrections made halfway through an explanation. What I accepted and what I disagreed with when AI showed me another view.

I want to retain that information in a form that comes before a short “list of my preferences.”

For example, if I were recorded only as “someone who values narrative consistency,” the fact that I liked Otonashi's contradiction might be overlooked.

“Someone who likes moving stories” would not explain which kinds of emotion I find convincing and which ways of producing them feel contrived.

This experience made me feel that such records could serve purposes beyond getting suggestions that fit my taste.

After giving AI enough of my perspective, I could have it reread the work from outside that perspective, connecting the impressions I had described with other scenes in the work.

I have not isolated how much the accumulation of my impressions contributed to this interpretation. Even so, it gave me hope that a record intended to help AI understand my thinking might also become material for expanding that thinking.

Even If the AI Changes, I Want to Keep the Source Record

I use ChatGPT now.

But I may not keep using only the same AI forever. I might want to try another one, or have several partners for discussing creative work.

If my impressions and the history of those conversations exist only inside the service I currently use, I would have to explain everything again from the beginning.

So I want to keep the emotional source record myself, as my own data.

The main things I want to preserve are:

• The impressions I first expressed.
• The criticism and alternative perspectives AI offered.
• What I accepted and what I corrected in response.

I would keep my own statements separate from AI's interpretations. I do not want an AI's summary of “the kind of person you are” to replace the emotional source record.

Summaries are useful for rereading and for giving a new AI an overview. But I also want to retain the original exchanges so that I can remake those summaries.

There is no guarantee that another AI will reproduce the same understanding or interpretation when given the same data. Still, it can be material for explaining again what has moved me.

I had been talking about my taste in order to make a novel.

Instead, I received a reading I would not have thought of myself.

That is why I want to save not only finished articles and settings, but also the impressions that led to them. Whatever AI I speak with next, I want to keep a starting point for expanding my thinking in my own hands.

---

The Dialogue Behind This Idea

The following passages are excerpts from the “Organizing Impressions for Creative Work” conversation, edited for this article. What I want to retain as the emotional source record is the original exchange before summarization; that is distinct from the edited version presented here. My impressions, AI's criticism and interpretations, and the exchange about the source of the interpretation are presented separately. These are readings by me and AI at the time, not an official explanation of the work.

My Impressions of Angel Beats! — An Edited Summary

I like this work a great deal. Still, I feel there are rough edges in the world's rules toward the end, and I wanted to see more of the process of getting to know each companion and helping them move on.

What I like is that people move forward by accepting painful experiences as part of their lives, rather than being saved by having their past erased.

Iwasawa had music. Even though Naoi lived in someone else's place, the person who suffered and made the effort was Naoi himself. Nor should Yuri's past simply be overwritten with happy memories. Those were the parts that appealed to me.

However, although the audience knows Naoi's past, I do not remember a scene in which Otonashi himself learns it in detail. I found it somewhat awkward, as storytelling, that he could then say something that reached its heart.

I also like the scene with Yui and Hinata. Hinata's appearance there felt a little abrupt, but I liked the emotional movement that allowed it to be received as “Perhaps there could have been a future in which I was loved too.”

And it is not all heavy material. There are plenty of ridiculous operations and jokes. Everyone seems to be having fun, which helps me keep watching. I liked that range, and the music too.

At the end, I think Otonashi no longer wants to part with Kanade. He cannot follow through on what he has told other people. Rather than dismissing that as a flaw, I took it as, “In the end, it isn't just about logic.”

Kanade then returns to him what he himself has believed. I valued the structure in which the protagonist's own beliefs are tested at the end.

My Request

Critique my impressions and give me another perspective.

Note: Edited from the request at the time.

AI's Criticism — Challenging My Reading

There are parts you fill in generously because you like the work, and others you may dismiss too quickly as a lack of screen time.

You see the work through the theme of “accepting the meaning of the past.” But Iwasawa's resolution is closer to something she reaches herself; Yui receives the possibility of a future from another person; Naoi's existence is affirmed; and Yuri lets go of her attachment. Their forms of acceptance differ. You may be making them fit a unified theme too neatly.

More time could develop the characters, but it might weaken the momentum of jumping from wild antics into the heart of someone's life and then rushing on again. The compression that seems like a flaw may also be part of the work's distinctive character.

Note: A summary of AI's response at the time.

Another Interpretation of Otonashi Offered by AI

From quite an early point, Otonashi is also “someone who creates a sense of purpose by saving other people.”

He loses his sister, pursues medicine, helps others at the accident site, and helps others move on in the afterlife.

His final proposal that he and Kanade stay and help the people who come later can therefore be read both as an excuse not to leave her and as another attempt to retreat into the role of rescuer.

Otonashi also needed to graduate from the role of rescuer.

Note: Excerpted and edited from AI's response at the time.

Other Perspectives It Offered

The jokes do not only ease the strain of the serious material. We can also read them as people who could not have an ordinary youth while alive finally experiencing it after death.

Seen that way, school life itself is a form of salvation, and the graduation ceremony is a ritual for finally bringing to an end the youth they could not have before.

Note: A summary of AI's response at the time.

I Asked Where the Interpretation Came From

I'd looked up various impressions before, but I hadn't seen this interpretation. Did you come up with it, or did you get it from somewhere?

Note: Irregularities in the speech transcription have been edited.

AI's Answer

That was an interpretation I put together on the spot using the impressions you had just given me. I didn't take it from a review site, and I hadn't searched the web at that point either.

However, I can't say that nobody in the world has had the same interpretation.

Note: Edited from AI's response at the time.

My Reaction and the Later Check

I simply thought it was a very good interpretation. I didn't have that perspective myself.

Could you look into whether that view has been recorded in a blog or an article?

Note: Edited from my statements at the time.

After this request, AI searched existing responses to the work. It reported that there were related elements and analyses, but that it was difficult to find an interpretation framed in this particular way.

What I have preserved here is a sequence of dialogue in which I described my impressions, received another reading, asked about its source, and had AI look into it.