Learn, compare, and apply ways to align and govern AI systems
Each of these approaches borrows from a tradition in law, political theory, or philosophy. See where it comes from, what it kept and what it left out, and run it yourself in Python.
The state of open-source alignment and governance is not meeting the moment. Let’s fix it.
Technical approaches to AI alignment and governance have increasingly drawn on traditions from moral philosophy, political theory, and legal theory, from constitutionalism and social choice to deliberative democracy and virtue ethics. For example, Constitutional AI encodes normative principles in an explicit, easily understandable natural-language artifact. Social-choice approaches aggregate human judgments according to formal axioms, and deliberative approaches derive principles through collective procedures. Yet, this transfer of concepts is often selective. Constitutional AI carries over the codification of rules, but not the frameworks to contest and revise them.
This encyclopedia maps these approaches into 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 each regime has a tradition behind it. For every regime, the encyclopedia lists the approaches placed in it and the methods behind each approach. It also reads each approach against the concepts of its tradition and records which of them it preserves, partially preserves, or omits.
A systematic search found 3,056 records that meet our inclusion criteria, and 1,033 of them are methods. The encyclopedia annotates 192 of these records in depth, grouped into 67 approaches. The regimes are unevenly populated. Constitutionalism alone contains 28% (290 / 1,033) of the methods, while Polycentric governance, Practical wisdom, and Contested custom contain only 6, 9, and 13, respectively. Most methods also treat normative disagreement as a problem to be resolved away: 58% (600 / 1,033) collapse conflict, and only 9% (98 / 1,033) keep it open to contest. The sparse regimes point to new directions for AI alignment and governance.
Each regime and approach links to working code in aiag, a Python library that implements the approaches. It loads any causal language model from the Hugging Face Hub, and its five packages follow the five loci of norms. Install it with pip install aiag. More about how the encyclopedia is built.