This may be contentious but the “includes are always and extends are sometimes” is a very common misconception which has almost taken over now as the de-facto meaning. Here’s a correct approach (in my view, and checked against Jacobson, Fowler, Larmen and 10 other references).
The key to Include and extend use case relationships is to realize that, common with the rest of UML, the dotted arrow between use cases is a dependency relationship. I’ll use the terms ‘base’, ‘included’ and ‘extending’ to refer to the use case roles.
A base use case is dependent on the included use case(s); without it/them the base use case is incomplete as the included use case(s) represent sub-sequences of the interaction that may happen always OR sometimes. (This is contrary to a popular misconception about this, what your use case suggests always happens in the main scenario and sometimes happens in alternate flows simply depends on what you choose as your main scenario; use cases can easily be restructured to represent a different flow as the main scenario and this should not matter).
In the best practice of one way dependency the base use case knows about (and refers to) the included use case, but the included use case shouldn’t ‘know’ about the base use case. This is why included use cases can be: a) base use cases in their own right and b) shared by a number of base use cases.
The extending use case is dependent on the base use case; it literally extends the behavior described by the base use case. The base use case should be a fully functional use case in its own right (‘include’s included of course) without the extending use case’s additional functionality.
Extending use cases can be used in several situations:
One important aspect to consider is that the extending use case can ‘insert’ behavior in several places in the base use case’s flow, not just in a single place as an included use case does. For this reason, it is highly unlikely that an extending use case will be suitable to extend more than one base use case.
As to dependency, the extending use case is dependent on the base use case and is again a one-way dependency, i.e. the base use case doesn’t need any reference to the extending use case in the sequence. That doesn’t mean you can’t demonstrate the extension points or add a x-ref to the extending use case elsewhere in the template, but the base use case must be able to work without the extending use case.
I hope I’ve shown that the common misconception of “includes are always, extends are sometimes” is either wrong or at best simplistic. This version actually makes more sense if you consider all the issues about the directionality of the arrows the misconception presents – in the correct model it’s just dependency and doesn’t potentially change if you refactor the use case contents.