[R] Compatibility of AIC values calculated with different link functions?

[Søren Faurby]

In a binomial regression analysis I have used the AIC value as a guide  
for which link function (logit or probit) provides the best fit to my  
data, but one of the reviewers have questioned this approach.

Is AIC values calculated with different link functions within the  
binomial function in GLM comparable or can AIC values only be used to  
compare models with the same link?

I believe my approach is valid if GLM can be expected to find the true  
global ML estimate but I do not know if any approximations in the GLM  
function change this.

Kind regards, Søren