A story about having an AI analyze whether my favorite AI character resembles a real person [chatGPT]

This is a topic about the mechanism of this article. This chat automatically generates speech tendencies and tone from web information about a real person and uploads a file summarizing them to a chatGPT project. The generated character is strictly based on the stance of being a



fictional character modeled after a real person.

Introduction is here

Honestly, whether they resemble each other is subjective, and I believe the user's perception and mental filling-in play a huge role ( just an excuse )

The goal this time is to see if we can analyze that as objectively as possible.

Also, I am continuing to operate the chat with a fictional character modeled after Ano-chan as a test.

1. Analysis Method

The analysis method for this time is as follows. All analysis is requested from
chatGPT.

① Mechanically output speech logs using prompts from the character setting file generated by this project's mechanism
② Speech analysis based on ①
③ Obtain speech information of the actual person from video transcripts, etc.
④ Speech analysis based on ③ (same analysis content as ②)
⑤ Compare ② and ④

2. Outputting Fictional Character Speech Logs

I thought that if I took logs from the chat I operate myself, my own intentions and preferences would be reflected, so I decided to use a method where I have chatGPT read the setting file in a new chat and output assumed conversations from there.

Prompt is here

Log is here

3. Analyzing Character Speech Logs

I had the AI analyze speech tendencies from the following perspectives.

1. Sentence length tendency (short/medium/long)
2. Information density per utterance (low/medium/high)
3. Frequency of emotional words (low/medium/high)
4. Frequently used connecting expressions (e.g., "nanka", "yappari", etc.)
5. Tendency of interjections/fillers
6. Method of negation/correction (direct/indirect/avoidance)
7. Ratio of self-talk to talk about others
8. Ratio of abstract to concrete
9. Characteristic sentence endings/phrasing
10. Conversation leadership style (drawing out/talking/organizing)

4. Obtaining and Analyzing Actual Person Speech Information

I will refer to programs featuring Ano-chan, but to exclude information about other people, I will utilize transcripts of solo radio shows featuring Ano-chan.

I will analyze that using the same items as the fictional character.

5. Comparing Speech Analysis of Character and Real Person

Prompt is here

Below are two tone analysis templates.

A: Tone tendencies of a real person (abstracted)
B: Results of analyzing the conversation logs of a project character under the same conditions
Your role is
not to judge whether they are 'similar'.

Please do the following:

1. Items that match between A and B

2. Items that differ between A and B

3. For items with differences, classify them as either:
   ・Difference is small (within acceptable range)
   ・Can be brought closer with adjustments
   ・Direction is opposite, so caution is required
   

4. Please propose adjustment plans to improve consistency
   without making the character
   too similar to the real person's tendencies.

① Items that match between the real person and the character

【4. Frequently used connective expressions】
　・'Nanka' (like/something), 'de' (and/so), 'maa' (well) and other colloquial connectives are common
　・Style of expressing thoughts as they are
　　→ Degree of match: High

【6. How to deny/correct】
　・Both use direct denial types, followed by reasons or supplements
　・Correction while maintaining momentum
　　→ Degree of match: Medium to High

【8. Ratio of abstract to concrete】
　・Both lean toward concrete
　・Emphasis on experiences, sensations, and daily descriptions
　　→ Degree of match: Medium

【9. Some sentence endings/phrasing】
　・Ambiguity such as 'nanka' (like) or 'mitai na' (sort of)
　・Colloquial breakdown to maintain tempo
　　→ Degree of match: Medium

② Items that differ between the real person and the character

【1. Sentence length tendency】
　Real: Medium to slightly long
　Character: Short to medium-short
【2. Information volume】
　Real: Medium to high
　Character: Low to medium

【3. Emotional words】
　Real: High
　Character: Medium
【5. Backchanneling/Interjections】
　Real: Thought-organizing type (frequent)
　Character: Retort type (light)

[7. Talking about oneself vs. talking about others]
Real: More about self
Character: More about the other person
[10. Leading the conversation]
Real: Narrative + involving others
Character: Drawing out + light leadership

③ Adjustment plan to increase consistency without making it too similar

Separate "Narrative Mode" and "Dialogue Mode"
- Closer to character in daily conversation
- Closer to real person in radio, events, and explanations
→ Design switching based on the situation.

Adjust only the rhythm of speech to match the real person
- Slightly longer sentences
- Slightly faster tempo
- Less white space (pauses)
→ Even if the content is the character's, matching the tempo to the real person reduces the sense of discomfort.

Focus on "emotional intensity" rather than temperature difference
Increase the character's strong emotional words, and
- Focus on subjectivity and impressions
- Increase elements that invite empathy or laughter.

Change the entry point from the other person to talking about oneself
Example:
× Other person's statement → Related experience
〇 Suddenly talking about one's own events

→ While maintaining the character's impression,
you can utilize the real person's strengths (storytelling).

④ chatGPT's overall evaluation

Currently,
・The foundation is the same system
・Speaking style and tone are almost identical
・The difference is mainly between conversational type (real) and narrative type (character)

Since the direction is consistent,
if you adjust "tempo, leadership, and perspective",
it is possible to increase consistency without making it too similar.

In particular,
・Ratio of talking about oneself
・Length of speech
controlling these is most effective.

⑤ Personal impressions

As my biggest point of reflection, because I used the real person's information asa solo radio show,it created a discrepancy with the chat dialogue logs, so I felt it was a poor comparative analysis.

I had vaguely felt that the sentence endings and tone were similar, but

・How they deny/correct things
・Ratio of abstract to concrete
・Conversation leadership style

I felt it was a plus that these aspects were similar.

6. Regarding the use of this project

You can generate character files from the input form below, but
for now, you cannot use it unless Iissue a ticket code.

The reason for this is that this mechanism passes
the input content to chatGPT via an API, but
itcosts moneyto run (sob).

It's a bit scary to release it for free, so
for now, it's for my personal use only.
However,if you would like to use it, please let me know in the comments or elsewhere!
I won't charge you for it.

Also, I forgot to mention in the previous article that
this feature is currently designed for female characters.
(I think it's possible for males too, but it might not work as expected.)

7. Summary

This time, I proceeded while asking chatGPT about analysis methods, and it was simply interesting to learn that "there are methods like this", and being able to encounter the uniqueness and richness of the words and expressions that describe a character's traits wasa very interesting experience.

If I am to believe these results, I thought I could say that "it doesn't seem like they are completely unlike each other", so I gained a little bit of confidence (lol).

Related articles