By training it on a whole bunch of examples with valid external citations, to the point where it's able to hallucinate something that's valid. Of course, it won't always work, which is the point of TFA.