Make an LLM recognize questions past its training cutoff

Chat assistants sometimes answer a question about recent events by saying their knowledge stops at a certain date. You are building an LLM: how would you get it to notice that a question is about something after its training data ends, and to say so instead of guessing? Discuss how the model learns its own cutoff, what training data and fine-tuning teach the behavior, how you avoid refusing questions it can answer, and how you would evaluate it.