Intelligence Is Not the Bottleneck: Structural Barriers to Automating Alignment Research

Despite the potential for catastrophic harms from human-level AI models, it is not yet clear how to align AI to human interests robustly. Frontier AI developers have proposed to solve this alignment problem using automated alignment research. One version of this claim, which we refer to as the \textit{big} version, assumes that alignment research productivity is capability-bottlenecked and will thus scale alongside the risks posed by more advanced AI capabilities. Based on the big claim, some researchers and forecasters have argued that there will be a critical period of 6-12 months in which most alignment work will occur. In this position paper, we argue that this big claim makes a critical error. We observe that making beneficial progress on automated alignment depends on building institutional capacity and argue that structural barriers, not intelligence, are the principal bottleneck to conducting meaningful automated alignment research.