Helen Toner: The Core Challenge of AI Alignment Lies in "Steerability"

Helen Toner proposes that "AI alignment" should be understood as "steerability": the real question is whether advanced AI systems can be reliably guided toward arbitrary goals, not whether they conform to someone's values. The term "alignment" carries an implicit reference frame, leading to conceptual confusion—nowadays in many circles, "alignment" is almost equated with content moderation techniques like RLHF, deviating from the steerability problem alignment originally referred to. She notes that one important reason for ChatGPT's success in November 2022 was OpenAI's adoption of RLHF derived from alignment research, causing outsiders to primarily use it to refer to such techniques.