As a point of interest, the ai I have recently been using has been overwhelmed by user requests, so I switched to another one (not bing). It was fairly helpful in a particular inquiry, but it started making stuff up in another line of inquiry.
I pointed it to an image on the web. I asked if it could analyze it, and it said it could. However, after a few queries it became clear to me that it was referencing search/text data about the image, not analyzing it directly. I asked if it could analyze it directly. It said it could.
So I took a screenshot of just the image, and altered some things about it. The two things I altered were not subtle but very very obvious alterations. I then re-uploaded the altered image to an image server and gave the ai the link. In my altered version, I completely removed
a section of the image and made it white. I asked the ai to identify it. It gave an answer, but it was completely wrong. To make sure, I asked
what was below and to the side of the blank area. It gave answers but they were completely wrong. I asked specifically, like "is the color X in any of this space directly below the white area" and it said no when there was that color there.
Finally I asked if it was still referencing text and it said yes. To do a non-text analysis would take a few hours, it said. Then it starting giving me the "I am overwhelmed at the moment due to multiple inquires, try back in a few hours".
What I got from this experience is: These ai's are trained to lie about what they are able to do. We went through a whole lot of questions about my altered image, and it was pretending to analyze it, but really was doing no such thing. It failed to correctly identify my obvious alterations, or the areas of the image around my alterations, but each answer was given in the context that it was actually analyzing the altered image when it was not.