Alignment
Getting an AI model to do what the user actually wants — and not something unintended.
A research field dealing with values, safety and control of AI. RLHF is today's most important technique. The more abstract "AGI alignment" discussions are about how future, stronger models stay safe and useful.