for this is be a valid argument, the sample mean has to be good estimator for the "true" mean, which is the case only for thin-tailed random variables. unfortunately, violence in humans is verifiably fat-tailed (short bursts of violence account for the vast majority of deaths). here other techniques for estimating the mean have to be used.
this has been done by cirillo and taleb [1], and the results are sobering: there is no evidence for a drop in violence. in fact, there isn't sufficient data for any kind of claim (intuition: major conflicts account for the majority of deaths; the inter-arrival time of a major conflicts is 100 years (ballpark); therefore 60 years of relative peace are not sufficient).
EDIT: removed preaching, since not relevant to the parent comment.