In the human domain, it's assumed that no generalization covers all cases, and that a single exception does not void a rule.
Not really on point, but memorable: even the proportion of heads to bodies in living humans isn't exactly 1:1.
For most generalizations about people , 80% accuracy is good, and 90% very good.
There's also a real secondary risk with these arguments that some genuine issue that e.g. affects 50% of XY's should be ignored because it also affects 5% of XX's aren't just wrong - IRL this is harmful in many ways.
One important example:
Much of the legislation intended to "make women safe from men" is ineffective, because it assumes that all men cause the specific harm being protected against, and is drafted on that basis.
In fact the vast majority of serious XY on XX crime is caused by a small minority ('Dark Triad", ASPD, etc (**)) who ignore the protective legislation, and continue with their bad behavior unchecked.
The same is true for XX on XY crime of course - the fact that XX's are less inclined to use physical violence doesn't mean they less likely to engage in serious anti-social behavior.
(**)
IIRC the total frequency in the population is low single-digit percent (perhaps 1 or 2%)