Presumably because it's similar enough to descriptions of problems and how they were solved that are in its training data.