Can the AUCs of ROCs for different sexes distinguish between sexual attitudes? I'm not a statistician but I've been reading all about ROC and AUC today and I"m getting closer but still don't quite get it.
If I have a model and I want to know if it's penalizing men or women differently, will running it on men and women and having a similar AUC automatically mean it's treating men and women the same?  Is it possible that say... false positives are higher with the women, and false negatives are higher with men (or something) and I still end up with the same AUC?  So the same model is basically dinging women and men in a different way.  Or perhaps the AUC is the same but once I set a threshold somewhere (and if I use the same threshold for women and men) then suddenly the results are worse for one gender than the other?
 A: If we just look at a graph, each point on the ROC-curve corresponds to the confusion matrix of classification at one threshold value i.e the false positives, false negatives, or equivalently sensitivities and specificities.

Is it possible that say... false positives are higher with the women, and false negatives are higher with men (or something) and I still end up with the same AUC?

Yes this is correct. See for example


Which will have the same AUC but have complete different set of sensitivities (=100-%false negatives) and specificities (=100-%false positives) for each threshold value.

If I have a model and I want to know if it's penalizing men or women differently, will running it on men and women and having a similar AUC automatically mean it's treating men and women the same?

No as discussed above. It's not part of the question but it's worth talking about the design of your study.
1) 
If you have a model that takes many variables as input and sex is one of them you could try feeding it the same input but only vary the sex-parameter. If it outputs different values what you want to see is how different. Here you have paired values as you have two outputs for each datapoint which you are varying the sex-input. 
There's multiple statistical tests for paired values (if simply tabulating and reporting the difference in sign isn't enough). I can give you more hints if wanted. If the location is significantly different you can  conclude that the model is literally discriminating based solely on sex.
2) If you only have the model outputs (score) for men and women then you want to compare how different the outputs are. Here you don't have paired values. 
AUC is a non-standard but excellent way of checking if the distribution of the score for men and women differs. You then calculate the AUC with sex as the class label. This is equivalent to reporting the rank-sum of the Mann-Whitney-U test for location. Read the wiki article and see how you can calculate the statistical significance. Location between sexes significantly differing doesn't warrant the conclusion that the model is discriminating as the inputs can differ between sex but you can conclude that the score differs.
A: I understand what you at hinting at, but, you would have to use slightly altered methodology to do what you want to do. The details are that you appear to be examining the same data in two different ways. When AUC is used to calculate probabilities of ROC curves, the underlying assumption is that the two ROC plots are being carried out on independent data sets. That is not what you are comparing. You are trying to compare two characteristics or classifiers on the same data set. That implies that the condition of independence is violated, so that the comparison is inappropriately of lower ability (or 'power') to distinguish between those characteristics (classifiers) than appropriate to the problem. That is, the classifiers are correlated which violates the independence assumption, or, if you wish, men and women are not only different but they also have characteristics in common.
So, what to do? One approach is to look at the ROC plots but do a test of difference of proportions as in this paper. Another approach is to do an ROC comparison for correlated curves. 
Finally, there are umpteen other statistical tests that might be applicable, e.g., ANOVA, and the question asked does not have a best answer without characterizing or typing the data for statistical properties like normality more exhaustively.
