A simple data race in C11/C++11 is undefined behavior, it doesn't have to result in any well-defined state that you might get by interleaving the execution of threads.
I think this has been on HN already: http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/why_undef.ht...
Essentially it has to be that way, or on some architectures every access to a potentially shared variable would have to be wrapped in fences by the compiler.
Strictly speaking, it could be implementation-defined behavior.
Ambiguity is quite frequently between several well-defined states. A significant difference between race conditions and natural language ambiguity is that in the latter we typically expect resolution to happen favorably when the interaction is cooperative - which is not a feature of race conditions in the general or typical case. I think it's still fair to say that this is not an essential feature of ambiguity, though, which is why I included it as an example. In the case of a race condition, the relevant pieces of the language do not uniquely define a single result - additional context may.
(the second sentence is a joke.)