I thought hallucination is when the model imagines something that isn't there at all. Not a classification issue where it misinterprets an object in the source data.

Sometimes it might be difficult to tell, but in the case of moon -> traffic light it seems clear.

Maybe I'm wrong about the terminology.

Isolating an object is baseline recognition. Any sort of semantic association is baseline interpretation. So that border is somewhere between "light source" and "traffic light".