The question of alignment—ensuring that AI systems pursue goals congruent with human intentions—has moved from a niche concern to a central research priority. The problem is multifaceted: technical alignment involves designing objective functions and training procedures that faithfully capture complex human preferences; conceptual alignment requires resolving philosophical uncertainties about what human values even are and how they might be aggregated. Reinforcement learning from human feedback has emerged as a practical technique for fine-tuning language models toward helpful and harmless behavior, but it remains a shallow approximation. Critics argue that it teaches models to simulate human approval rather than to internalize ethical reasoning. More sophisticated approaches, including debate, recursive reward modeling, and constitutional AI, attempt to address this limitation by structuring the learning process to incentivize truthfulness and corrigibility.