The ability of an LLM to produce something not in its training data set.