It is non-deterministic in the sense that the program itself and its inputs are not sufficient to determine the program behavior, as the amb operator creates a set of "possible worlds" in which the same expression has different values. Of course to implement it on a traditional computer, one has to simulate this behavior and decide on some order of evaluation. In SICP, where this is discussed in more depth, one of the exercises asks the reader to modify amb to return a random choice instead of just the subsequent one and this expands the range of problems one can apply this technique to:

http://mitpress.mit.edu/sicp/full-text/book/book-Z-H-28.html... http://mitpress.mit.edu/sicp/full-text/book/book-Z-H-28.html...