Generative AI struggles to understand 'a cat that eats a red-headed fish'

There is a sentence that once became a hot topic on Twitter: "a cat that eats a red-headed fish". According to the creator, Akihiro Nakamura, this sentence can be interpreted in five different ways. I have used generative AI to illustrate each of these five interpretations and will present them here. Getting the AI to understand this Japanese was quite difficult.

About the five interpretations

1. (A red-headed cat) is eating a fish

2. A cat is eating (a red-headed fish)

3. A head is a cat that eats (a red fish)

4. A being with the appearance of (a cat's head) is eating (a red fish)

5. A being with the appearance of (a red cat's head) is eating a fish

Usually, people think of 2, but 1 is also within the realm of common sense as it is an image of a cat eating a fish. However, the images for 3 through 5 contain elements of pure fantasy or horror. I think anyone who imagines these from the start has a very unique way of thinking. As for 3, it is a monster whose head is eating a fish, and in fact, it was the hardest of the five to illustrate with AI. Furthermore, for details on "a cat that eats a red-headed fish," please refer to Mr. Nakamura's X, where it is explained in an easy-to-understand way.

I tried generating images with DALL-E

1. (A red-headed cat) eats a fish

Since it won't generate it with the original text alone, I will instruct it by enclosing the subject in parentheses as shown in the heading.

2. A cat eats (a red-headed fish)

Even with the original instructions without any special thought, the intended image was generated. If I had to be greedy, it would have been better if only the fish's head were red, but let's be satisfied with this.

3. A cat whose head is eating (a red fish)

This was the most difficult one. It seems even the AI cannot handle this concept, and although I tried various things, I couldn't get an image that felt right until the end. Perhaps DALL-E restricts grotesque expressions. Even so, some ugly monsters were generated, so I will post a safe image here.

4. A being with the appearance of (a cat's head) is eating (a red fish)

When generating an image from this interpretation, you cannot get the image you want unless you clarify the subject hidden in the original sentence, which is a human with a cat's head, and input that.
If you input that it is a human with a cat's head, you can generate the image you want relatively easily.

5. Someone with the appearance of (a red cat's head) is eating a fish

Just like with 4, by clarifying the subject hidden in the original sentence, I was able to generate the image I wanted relatively easily.

Summary

This example clearly shows that Japanese is not well-suited for AI prompts (instructions) because it is a language that omits subjects and conveys images through context. That is precisely why this case illustrates the importance of making prompts clear and concise. In that respect, English can be said to be a language well-suited for AI prompts. I think the technique of translating Japanese prompts into English before giving instructions is effective for that reason.

Bonus

To tell the truth, when generating these images of cats?, I used the notation of LISP, a classical programming language.
The prompts used to generate the images in the title were
"(human(head(cat(eat(fish red)))))" and "(human(head(cat(red(eat(fish))))))", which I input into DALL-E to generate them.
Since AI understands programming languages more accurately than natural languages, it is effective to borrow notation from various programming languages.
It is difficult for me to talk about LISP, which has enthusiastic followers, so if you are interested, please refer to Wikipedia or videos that explain it in an easy-to-understand way.

I will continue to post articles mainly on digital-related topics, including information related to generative AI. If you like, I would appreciate it if you could like or follow me.