Comparing the AICs roughly corresponds to a LR-test with a level of
about .16 (Greven S, Kneib T(2009),
http://www.bepress.com/jhubiostat/paper202/), and the standard
likelihood ratio tests or F-tests will give you a (very) conservative
test (as you can see from Ben Bolker's GLMM -Wiki).
Package RLRsim provides an exact restricted likelihood ratio test for
testing single random effects in linear mixed models. It only works if
the random effect you're testing is uncorrelated with the remaining
random effects in the model, though.
HTH,
Fabian