AI models are supposed to spot when they should stop responding to sensitive topics. I saw how and why those plans can fail.