Cognitive Bottlenecks Improve AI Alignment with Human Moral Judgment

As AI becomes increasingly integrated into daily life and high-stakes decision-making, aligning it with human moral judgments is a challenge of profound importance. People often mistrust AI’s moral decisions, partly because its opaque processes differ from human reasoning. We introduce “cognitive bottlenecks” that guide large language models (LLMs) to base moral judgments on a small set of psychologically meaningful features, such as perceived harm, agent intention, and victim vulnerability. Across multiple LLMs and moral frameworks, bottlenecks improved alignment with human moral judgments. Participants viewed bottleneck systems as more moral and trustworthy than systems without bottlenecks, and preferred bottleneck-generated explanations over those produced through standard chain-of-thought prompting. Bottlenecks can be implemented through prompting alone, without model retraining. These findings demonstrate the benefits of integrating psychological theory into AI and offer a scalable approach to improving AI-human alignment.