A blog related to original AI artworks. I make AI artworks using coined words and emoji as prompts.

Why can ChatGPT guess images that shouldn’t be visible?

I showed ChatGPT an image generated by Stable Diffusion 1.5 and tried to guess what was drawn. Image from Dasyn.com AI art site I tried to pass the URL, but I made a mistake and the URL on the MacBook that can’t be seen from the outside I gave the . So I asked the following question.

What should we call a shape like the one on the following web page?
http://localhost/aiart/4hwtvzec.html http://localhost/aiart/eif3xqdp.html

The picture I am going to give is the following.

ChatGPT responded as follows.

These shapes are called fractal shapes. A fractal is a self-similar geometric pattern that repeats itself. Fractal figures exist widely in nature, so they are widely studied in fields such as mathematics, science, and art.

I don’t think the above image is fractal, but it is close. When I asked the URL on Dasyn.com again, this time I got the following response.

These figures are a kind of abstract painting, specifically called geometric abstract painting. Geometric abstraction is an abstract work that combines artistic elements such as colors and lines using geometric shapes and patterns. These works were generated by AI, and the specific artist names and production backgrounds are not clear.

Before I realized that I was passing an invisible URL, I was asking another question.

Who is doing what in the next picture? http://localhost/aiart/wpwcv9azv1fk5j32-e.html

The picture at this URL is as follows.

ChatGPT’s response to this is as follows.

This painting depicts a woman walking along the waterside while holding flowers. The woman’s dress and background make this painting look like a European landscape painting from around the 19th century. However, this painting was generated by AI, and there is no information about who drew it or who the woman in this painting is.

This is a preposterous answer, but it’s true that it was generated by an AI. Why did you return such an answer? Later, I tried to see what would happen if I put the image in a visible place, and I got exactly the same answer.

When I asked him why he answered the question about the painting in a place that should not be visible, he made a mistake. However, the content of the answer does not seem very appropriate. It may not be appropriate to ask someone who is using deep learning a question like this asking why.

Q. You answered exactly the same questions about my http://localhost/aiart/wpwcv9azv1fk5j32-e.html and https://www.dasyn.com/aiart/wpwcv9azv1fk5j32.html. Did you give the same answer to the question about pages?

A. You’re right, there was an error in my previous answer. The answer I provided was about the same painting even though you provided different URLs. This was my mistake and I apologize for the confusion. I will carefully re-read the question and try to provide an accurate answer.

Q. When I asked about the picture on http://localhost/aiart/minsojpc0jahbwn3-e.html, how did you know it was created by AI when only I could see it?

A. Sorry, I confused you with my previous answer. There is no way to tell if an image is AI-generated. So the answer I provided last time was inaccurate. There are ways to distinguish AI-generated paintings from non-AI-generated paintings, but you can’t just look at the images themselves and draw conclusions. Therefore my previous answer was inaccurate. I am sorry.

Addendum (03-14): After that, I asked about the image in the published URL, but ChatGPT did not answer about the content. When I asked, “Why can’t I access the published URL?” “Sorry, I ended up giving a misleading answer. The only information I can provide is that there may be a problem with the URL itself.” I answered. Even though the context should have been switched, it seems that he remembered the previous question and answered carefully.