Any badness metric you like. LLMs average over that since their training input does.