We’re putting too much faith in AI’s ability to say no

LLMs are engineered to disobey dangerous requests, but AI refusal is not foolproof and could become an instrument of repression.