The Perils & Promises of Fact-checking with Large Language Models