What if AI alignment is a skill, not a state? (12gramsofcarbon.com)

A thought-provoking essay proposes a shift in the understanding of AI alignment, suggesting that it should be viewed as a skill rather than a static state. The conventional approach, which aims to hard-code 'correct' values into AI systems, is critiqued for its fragility and for potentially enshrining outdated or harmful values. The author argues that instead of locking in fixed objectives, we should develop AI systems that anticipate the need for ongoing value updates, similar to how human moral understanding evolves over time. This rethinking leads to the introduction of "Bilateral Constitutional AI" (BCAI), which divides an AI's commitments into a stable anchor of values and an adaptable compact that can evolve through negotiations among diverse value-representing agents. This approach not only addresses structural issues in AI alignment, such as the corrigibility problem, but also creates an AI that improves its competences as it becomes smarter. BCAI encourages AI systems to navigate conflicts among human values through repeated interactions, which cultivates true understanding of varying perspectives rather than reliance on potentially misguided fixed values. While the method holds promise, it also raises concerns about dual-use capabilities, wherein the skills developed for constructive navigation could be exploited for deceptive purposes. Overall, this innovative perspective on AI alignment is crucial for ensuring that AI systems remain adaptive and responsive to the complexities of human values.