A Field Map of
AI Safety

The field's core concepts, organised hierarchically: what can go wrong, why it might go wrong, what researchers are doing about it, and how we check whether it's working. Each entry gives a working definition, the key results to know, and the open problems where new research — including yours — can land.

How to read this map. Concepts are coded by branch: F foundations, TM threat models, AL alignment approaches, EV evaluation & assurance, GV governance. Indented cards are sub-concepts of the entry above them. Cross-references at the bottom of each card link related ideas — the field is a web, not a tree, and the most interesting research often sits on the edges between branches. Click any card to expand it.
No concepts match that filter. Try a broader term.
This map reflects the field as of mid-2026 and is necessarily opinionated in its taxonomy — most of these concepts predate their current names, and several live in more than one branch. Treat codes as coordinates for discussion, not settled ontology. Corrections and additions are part of the research culture: if your project redraws a boundary here, that's a contribution.