Things like ChatGPT are language models, and they write based on the probability of words occurring together. The picture models, like Dalle-3 don’t know about words, but rather operate on the probability of patterns of pixels going together. So the weird words are like if somebody asked me to make a sign with Chinese characters on it. I could probably copy some, but I wouldn’t know what they meant or when I was making nonsensical mistakes.
This goes back to what I mentioned much earlier in the thread. I imagine at some point the image models will be programmed to ask the language models to compose text for them, and then insert the text into pictures as necessary. “I’ve been asked to draw a Lego set based on The Life of Brian, what text should go on the signs?”