People arrive at AI Safety from very different places: machine learning, law, philosophy, security, public policy. Each brings a different picture of what the field is about. This article offers a simple map.

Three questions

Most work in the field can be placed under one of three questions.

  1. Can we get AI systems to do what we intend? This is the domain of alignment and specification: training methods, feedback, and the ways objectives go wrong.
  2. Can we tell whether they do? Evaluations, interpretability and robustness research all aim to measure and understand what a system will actually do, especially in situations we did not test.
  3. Can our institutions keep up? Governance, law and policy ask who is accountable, which rules apply, and how decisions about deployment get made.

Where security fits

Security and misuse research sits across all three. A perfectly aligned model can still be attacked, and a well-governed organisation can still deploy a system that is easy to misuse. Threat modelling — thinking systematically about who could cause harm and how — is a skill that every part of the field relies on.

Where to start

If you are new, read one accessible overview, then one paper from the area closest to your background. The AI Safety Foundations pathway on this site is designed for exactly that, and the Knowledge Library lets you filter by topic and difficulty.

A field map is a starting point, not a hierarchy. The boundaries between these areas are porous, and much of the most interesting work happens where they meet.