There may be multiple pairwise alignments that have the same optimal score with respect to gap penalty and match scoring. The function just returns the first it finds during the traceback of the dynamic programming matrix rather than returning all of them.

From the following link, several paragraphs down, you will find this quote:

"Our variable alignments now contains a list of alignments (at least one) which have the same optimal score for the given conditions."

http://biopython.org/DIST/docs/tutorial/Tutorial.html#htoc86

More concretely, you can check the source code. The function you are interested in is _recover_alignments, where the parameter one_alignment_only is put to use.

http://biopython.org/DIST/docs/api/Bio.pairwise2-pysrc.html