AI language models can tell real events from impossible ones, hinting at emerging common sense, according to a new study.