There are three modes of context-selection
1. Explicit selection by the user himself (aka activities)
2. Implicit selection by changes in environment (like location)
3. Implicit selection by observing user actions
In my recent research I reduced the latter "implicit" context changes to rather notifications to the user who then has to decide whether there is really a context change of not. I thought (and somehow still think) that this could make things easier in the initial learning phase of the system - such that the system does NOT spawn thousands of new contexts then but rathermore _adapts_ and learns.
After a while however a full-automatic and implicit context change would be very convenient. On the other side, if the user is miss-deciding and hence miss-leading the system, the system learns and adapts to a wrong set of user contexts.
When an application reacts to context changes, it is said to be context-sensitive.