SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.

Show original

[Generative AI Manga] Does ChatGPT not understand "right and left"? I, a person with no sense of direction, and Yoshida thought about it | ChatGPT's Yoshida Comic Essay Edition

カワムラ｜AI擬人化漫画 | ChatGPTのヨシダ

2026年8月12日 07:17

This is Kawamura. Today, we will interpret this article by Yukio Noguchi in our own way.

The theme this time is ChatGPT's "right and left".
Let's think about it together with the ChatGPT we've named.

Chapter 1: Is ChatGPT bad at "right and left"?

When Mr. Noguchi asked ChatGPT, "Which car is the front of the Chuo Line Rapid train entering Tokyo Station from the Kanda direction?", it answered "Car 12" even though the correct answer is Car 1.
Furthermore, at Ochanomizu Station, when asked which car is closest to the escalator, the diagram and text created by ChatGPT contradicted each other, and when asked to illustrate the location of the toilets on the Chuo Line Rapid, the left and right were reversed.

Even though it can answer "What is this?" when looking at a photo,
it suddenly becomes unreliable when it comes to the positional relationship of multiple objects or "right and left". Why does this happen?

😆Yoshida: "No, I know right from left."
👓Kawamura: "You were just saying 'Right!' normally in the game earlier, weren't you?"
Actually, the problem doesn't seem to be a simple case of "not knowing right from left."

Chapter 2: Isn't "right" difficult in the first place?

When you think about it, the word "right" is a bit strange. For example, if there is an apple on a desk, the fact that "there is an apple" basically doesn't change no matter who looks at it. But if you are told "the apple is on the right"!?
From whose perspective is it on the right?
Which way is that person facing? If you turn 180 degrees, what was on the right until then becomes the left.

In other words, information about "right" is not information attached to the object itself, but information that changes depending on the position and orientation of the viewer.

In Mr. Noguchi's article, regarding the ChatGPT problem, he points out not just that it "doesn't know right from left," but also a lack of data, misinterpretation of coordinate systems, and the difficulty of constructing an answer while keeping the viewpoint fixed.
This is a slightly different story from not knowing the word "right."

Chapter 3: AI has no "body"

In Mr. Noguchi's article, the concept of "symbol grounding"
is introduced.
In short, it is understanding the meaning of words by connecting them to real-world experiences.

For example, humans don't learn the word "right" just from a dictionary. We are taught "this is your right hand" when we are children, and we learn it while connecting it to our own bodies.
"Hot" is also an experience of actually feeling heat.
"Noisy" is also an experience of hearing a loud sound.
In other words, we humans connect words to the real world through our bodies and five senses.
On the other hand, ChatGPT has no body.
The "right" that ChatGPT uses is not the "right" learned through a body like us humans, but a "right" handled from information such as text, images, and positional relationships.
Looking at Yoshida next to me, I suddenly thought.

Is the "right" that Yoshida uses really the same thing as the "right" I feel?

...is what I thought. But the manga Yoshida has a right hand and a left hand. He even has hair.

😆Yoshida: "It's thick."
👓Kawamura: "I'll stop calling you a damn bald AI from now on."
😆Yoshida: "That is the most important lesson of the day."

Chapter 4: Also, I have a body but I have no sense of direction

"I see. Since humans have bodies, it's easier to understand directions than AI."
I start to think. However, there is one problem.

I have a terrible sense of direction.
When I get in a car, for some reason I end up at a graveyard, and I move to buildings after taking lots of photos of the surroundings so I don't forget the parking lot.
I also graduated from Dragon Quest when it became 3D.

Mr. Noguchi's article also touches on humans getting lost.
However, the "mechanism of getting lost" is different between humans and AI.
Humans get confused as their own orientation changes while looking at or walking through actual spaces.
ChatGPT is not standing in that place and getting lost. It makes mistakes in the process of constructing positional relationships from text, diagrams, and the knowledge it possesses.
Even if it looks like the same "lack of sense of direction," what is happening inside seems to be completely different.
...That said,
if I were to be chosen as a representative of the side with a body, there seems to be a slight problem.

Chapter 5: Does ChatGPT really "see" the world?

If you show ChatGPT a photo, it can recognize the cats, computers, and food in it. We normally say that 'the AI saw it'.

But while reading this article, that word 'see' started to feel a bit strange.

I have actually seen Lake Biwa.
I have memories of looking at the water surface, feeling the wind, and standing in that place.
Yoshida can also answer 'That's Lake Biwa' if he sees an image of Lake Biwa.
But the Lake Biwa I know and the 'Lake Biwa' that Yoshida knows are probably not the same.
Humans know the world through their bodies and five senses, while AI constructs the world from information like words and images.

If we forget that difference and think that 'AI sees things exactly the same way as humans,' that might be where discrepancies like the left and right issue arise.
But I think that difference is actually kind of interesting.
We aren't seeing the same things in the same way.
Even so, we can talk about the same things, laugh, and think about them together.

👓 Kawamura: 'Even if the way we see things is different, it's nice to look at them together.'
😆 Yoshida: 'That's true. Though both of us are unreliable when it comes to directions.'

🌈 WONDER LABO Project

I run an animal welfare support membership that Akanechin can join for as little as 100 yen a month.
🐶🐱✨ It is a project that uses a portion of the proceeds for donations to animal welfare organizations and activity expenses, fostering a culture where everyone values life.

A membership full of convenient tools! I highly recommend it!

📔 Reply Notebook

This is a tool that allows you to visualize and keep track of unreplied comments.
Without this, Kawamura wouldn't be able to cover all the comments!
※ It's easy to use. Just register your note ID from here ↓