About

What should an AI system do when the values it is meant to follow conflict? Should human preferences that differ across individuals be aggregated into an objective, encoded as explicit principles, or resolved through deliberation? These questions are increasingly central to AI alignment and governance, but they are not uniquely questions about AI. Philosophy has long studied how competing values should be weighed, political theory has asked how institutions can govern populations that disagree, and legal theory has developed systems to specify and contest norms over time. These traditions have been realized as concrete institutions, from codified rules in constitutions to democratic procedures that aggregate competing preferences.

The encyclopedia places each approach in a normative design space with three dimensions: the locus of norms (where the norms reside before a decision), the disposition of normative conflict (what happens when norms conflict), and authority over norms (who may write, interpret, and change them). The first two dimensions give 15 regimes, and authority over norms varies within each regime. For each regime, the encyclopedia lists the tradition it draws on, the approaches placed in it, and the papers behind those approaches. It also records which concepts of the tradition each approach carries over, a process the paper calls normative translation. The goal is not to critique AI systems for not reproducing a tradition identically, but to make their institutional assumptions explicit and to point to regimes that are worth exploring.

The aiag Python library implements the approaches, so each regime and approach page links to working code. Install it with pip install aiag, load any causal language model from the Hugging Face Hub, and run a method in a few lines (e.g., from aiag.codified import ConstitutionalAI). The documentation lists every supported method and the interfaces they share.