In building a conversational AI, should we prioritize natural language understanding over precise, literal responses?