It's because it's a statistical model that guesses the next word. You asked it to read a document and formulate a response. So, based on responses in the dataset to the documents in the dataset, it chose the words and phrases which were most likely to be in a response.

Statistically, the most likely result the machine could come up with, was that you had used the wrong word. Factually, you did not.