This is interesting in terms of information theory. When you compress information you are basically figuring out the pattern behind something, and encoding that pattern instead of the entire image. For example, PNG which assumes that pixels are more likely to be the same or similar to the pixels above or to the left of them. So you only have to encode how much difference each pixel has to it's neighbors, rather than the entire value of every pixel.

If you were to randomly corrupt a highly redundant format, like a bitmap, it would just change a few pixels. On more compressed formats like JPEG it seems to affect the entire image, and in very specific ways (mainly the color of every block of pixels after the point that was corrupted.)

If you corrupted a perfectly compressed image, it would give you a different image entirely, possibly of something very similar. I.e. if you had a image format very good at compressing faces, corrupting it would result in a different face entirely, not randomly misplaced pixels or colors. And the face would be similar to the original, maybe with stuff like a different nose type, or an extra freckle.

The corruption is revealing what kinds of assumptions it is making about the content.