AI Alignment & Governance Encyclopedia

The Home of Open, Accountable Alignment & Governance for AI Models

There are many ways to make a model behave as its creators and users intend. This is the unified resource for understanding and applying them.

15 regimes·67 technical approaches·239 works indexed

What this is

The best-known way of giving an AI system binding norms is a written constitution: a document settled before deployment, applied to every case, and resolved into a single answer whenever its clauses pull against each other. That arrangement is one square of a much larger board, and it is not obviously the right one.

This encyclopedia maps the board. It crosses two questions — where a system’s binding norms live, and what the system does when those norms conflict — into 15 distinct governance regimes, each with a tradition in law, political theory or moral philosophy standing behind it. For every regime it records the approaches that place in it, the working methods behind each approach, and a concept-by-concept reading of how much of the borrowed tradition the built work actually carries. Borrowing Hart’s vocabulary and delivering a fifth of Hart is a finding, and the map is built to show it.

It currently indexes 239 works across 67 approaches. 93 of those works are methods — a mechanism given in enough detail to reimplement and shown working by its own authors — and only those count anywhere on the map. The empty space matters as much as the full: 3 of the 15 regimes have at most two working methods in them, and one has none at all. More about how it is built.

Ways in

The two axes

The rows

Locus of authority

Before any particular case comes up, where does the binding content actually sit? Not who wrote it, and not whether it is any good — just where it is, and therefore what you would have to change to change the system’s norms.

Codified
A specification settled before deployment, complete in principle before any case arises.
Case-accumulated
A growing body of decided cases. Nothing is settled in advance; what to do is whatever the record has held.
Customary
Regularities of practice, never written down and never adjudicated.
Dispositional
Trained disposition rather than any external artifact. There is nothing to look up.
Procedural
No norms persist between decisions; a procedure — a vote, a deliberation, a debate — produces the answer each time.

The columns

Disposition of conflict

Two of the system’s norms point opposite ways, or two of the constituencies behind them do. What does the system do with that disagreement? This is the axis the field is least even on, and the one that decides whether anybody can argue back.

Collapse
The disagreement is aggregated, averaged or adjudicated away, and the output carries no trace of it.
Contest
Either the output carries the conflict, or the decision stays open to challenge through a standing channel.
Partition
Several normative orders are kept intact, and a rule assigns which one governs the case.

Crossing the two gives fifteen regimes — the grid is that crossing, and the configuration finder asks you these same two questions and tells you what has been built where your answers land.