I was pondering about how you could use ChatGPT to help visually impaired people understand what images look like.
So I thought I’d ask it to tell me about some non-famous paintings.
The first painting ChatGPT described to me was Mary Cassatt’s The Rosebush. I’ve never heard of the artist before, so it’s a good start.
A solid description, and I have an image in my mind’s eye. I turn to Google to compare it to the image, and… it doesn’t exist. ChatGPT literally made up a painting for me. A plausible one given Mary Cassatt’s other works, but completely fictional.
Next try, I find the name of a painting, Drinka Pinta Milk (1962), and turn to ChatGPT again.
I’m pretty sure Banksy was not around in 1962. Maybe there’s two paintings with the same name?
No, ChatGPT. Banksy did not paint an image called Drinka Pinta Milk. Once again you’ve completely fabricated something.
One last try, this time give it the artist’s name (Derek Boshier) and date (1962). It describes a picture in great detail, I’ve got a fantastic image in my head:
Got an image in your head?
You can see the actual painting here.
I wish I had the confidence of ChatGPT. Although I’m not sure I could be that brazen.
On a serious note, I find it incredibly alarming at how much is going to be taken at face value, quoted and used to inform real decisions.