Cosma Shalizi has a piece about how if you aggregate any similar set of tests, even if they're meaningless, even if you literally generate them at random, you will get a "g"; it's like pointing out that a nonsingular matrix has an inverse.
Interesting, but also irrelevant - because the sorts of tests in question certainly are meaningful; it's not obvious a priori that such a correlation should be expected between them; and the extracted correlate has demonstrable predictive power.
http://bactra.org/weblog/523.html