Field Guide / Seeing the system
Sociotechnical systems
A sociotechnical system is the whole working arrangement an AI tool lives inside: the model, the people and agents who use it, the records it reads and writes, and the institutional rules that govern all of them. Safety belongs to that arrangement, not to any single part.
Most conversations about AI safety focus on the model: how often it is wrong, what it refuses to do, how it scored on an evaluation. Those questions matter, but they describe one component. The model is one node inside a larger system that includes the people who act on its outputs, the case records and databases it draws from and feeds, and the institution — with its deadlines, caseloads, and incentives — wrapped around everything.
That larger system has its own behavior. An accurate model inside a rushed, under-staffed office can produce worse outcomes than a mediocre model inside a well-governed one, because what reaches a decision is never the raw model output — it is the output as read, trusted, recorded, and acted on by people under pressure.
This is why checklists and model cards, useful as they are, cannot carry responsible AI on their own. They describe the component. Governance has to address the arrangement: who checks what, which records can be written by whom, what happens when the reviewer is overloaded, and who has the authority to change any of it.
The Field Guide builds on this view throughout: the system map names the parts, the propagation loops describe how errors move between them, and the governing-adaptively concept describes how institutions can steer the whole.
This page is conceptual framing — a way of seeing, not an empirical claim. Documented real-world events appear in the Domain Atlas with citations; testable versions of these ideas live in the PAN Lab.