Because the software doesn’t understand the concept of anything. It’s been fed a bunch images of keyboards and programmed to output an average of the images. It has no concept of a pattern or a keyboard in which to conform to something recognizable.

Kinda depends on the generator. I saw one the other day that was of a guy in a bedroom on an oil rig. It had an HP Chromebook on the desk. As in I was able to reverse image search it to find the brand of computer it was. The entire image was AI generated. There were a few details off, a few weird things about the image that didn't make sense on closer inspection but by in large, it looked like a regular old selfie.

Depends on the model. Many of the cheap and light image generation models are trained to be good at the most obvious stuff so they can run on weaker hardware. They can generate perfectly realistic and consistent faces (and certain other body parts that are... in high demand) because that's what the human eye instinctively focuses on first, but will struggle to generate a wooden chair that isn't a weird M.C. Escher mess. You can run those models on a consumer GPU to generate multiple images per minute.

Not all models are trained with those limitations. Some can get scary accurate and consistent.

Right? It's like they get the idea that a stand in for the concept of a "pattern" should go here, but no concept of what a pattern is.

I think the issue is the same on the text generation side, but harder to notice because it's not an image.