No, it does not know what a car is. It cannot abstract the idea of "car" from it's training database. It can't even abstract the idea of "object" from its database.
This is why it's so terrible with hands. I has no abstract idea of what a hand it, what it does, what it looks like. It can only concatenate several millions of images tagged with the word "hand" and calculate an average of what lines, shapes and coloration correspond to that tag. But because hands are so mobile, fluid and expressive, the examples do not do much to constrain the image generator.
Likewise, it has no idea of what a "car" is. It has a list of compositional elements, shapes, lines, curves, etc that correspond to the tag "car". The image generator has no concept of what a car is, or even what a car looks like, because concept is itself outside the programming.
This is a category error.
And AIs can't observe.