Pergunta de entrevista da empresa Anthropic

How would you design a language model that minimizes harmful outputs while still being useful and expressive?