Nothing in a large model holds a capability the way a file holds a value. Parameters take part in distributed representations and non-linear relationships with one another, and at sufficient scale those relationships produce higher-order structure — structure that produces behavior which cannot readily be attributed to any individual parameter. The chain runs from parameters, to the interactions among them, to higher-order structure and dynamics, and from there to emergent capabilities and emergent risks.
The pattern is ordinary in complex systems, which is what makes it credible rather than mystical. No single neuron contains a thought. No individual molecule is wet, though water is. No one ant carries the colony's foraging strategy, yet the colony has one. In each case the property belongs to the interactions rather than to any part, and it appears only once there are enough parts, richly enough connected, for those interactions to constitute a system in their own right.
This is not the same claim as parameter scaling, and the difference is the whole point. Scaling adds parameters. Parameter emergence is what those parameters begin to do to one another once there are enough of them, at sufficient complexity, to behave as a system. That is why capabilities nobody specified and failure modes nobody designed tend to arrive together: they are two faces of the same transition, not a benefit and a separate, unrelated cost.
For governance the consequence is sharp. If a behavior is a property of a configuration rather than of a component, then examining components will not find it, and a scaling curve cannot be counted on to say when it appears. Evaluation that decomposes the system into parts and checks each part answers a necessary question — but a different one, because it describes the components while the behavior lives in their arrangement.
The same logic holds one level up. A deployed AI system is also a configuration of interacting parts, and the sociotechnical-systems concept is this argument at deployment scale: behavior that belongs to the arrangement, not to the model measured alone. Parameter emergence is one reason that view is more than a broader perspective — and the system map names the parts of the larger configuration it applies to.
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.