It doesn’t copy paste. These larger models seem to have a context capable of quoting, but it still doesn’t copy paste.
It essentially reads incoming data and builds a huge multidimensional map over the words and the relationships between them. Then when given an input it traverses this map to figure out the most likely destination.
Depending on how you tune it the outcome will differ from one occasion to the next.
It’s a gross oversimplification but a closer approximation than “copy paste.”