You're not wrong here about how a dummy designed to "the average male" proportions" is going to exclude men whose proportions exist in the statistical extremes (and likewise for women), but a phrase that comes to mind is "All models are wrong, some are useful". Whenever we are making a model for gathering and analysing data, it's because the real phenomena we're studying is too complex to be able to effectively analyse without a model. Even if we had a wide array of anatomically correct crash test dummies for many different body types, it would still be a huge simplification of reality. A huge part of research is about trying to always be mindful of this tension, and to be constantly evaluating whether our model is a good enough for reality.

The stats for injury rates in women indicated that no, our model was failing pretty significantly in this area, so we designed different dummies, effectively updating our model. There will be more research that looks at other kinds of variation between people, and that will mean trying to account for extremes while not overcomplicating our model.