Text generation, when fed the data it needs, is a fully general task API. You would submit something like the following request to GPT-3.

  Wikipedia sentence: In 1735 a volcano on the island of Foobar erupted.

  Original article content: In the history of volcanic eruptions there have been many different notable eruptions. One occurred in the year 1601 on the island of Foobar. Another occurred 134 years later on the same island.

  Does the citation match: Yes.

  Wikipedia sentence: Bar Foobar was a prolific American poet.

  Original article content: In recent years, the number of athletes who have been added to the American Sports Hall of Fame has increased dramatically. In fact, the Sports Hall of Fame just added its 1000th new member, Bar Foobar, on June 15, 2014.

  Does the citation match: No.

  Wikipedia sentence: The leader of the Foolandian revolution was Bob Jones Sr.

  Original article content: The Foolandian revolution remains one of the hallmark events of the 20th century. It was spearheaded by Bob Jones Sr. who today has a Foolandian holiday named after him.

  Does the citation match:

GPT-3 then fills in the last part (adding a Yes or No after "Does the citation match:") and you'd vary the final "Wikipedia sentence" and "Original article content." You'd probably flesh out the examples a little and add a few more of them, but that's the general idea. GPT-3 is very very good at delivering results in the right format (given this input, I would be very surprised if GPT-3 ever delivered something other than "Yes" or "No"). Whether it would be clever enough to do the matching well is a separate story, but I think it'd actually do a pretty good job since this is pretty up its alley.