I try to get around this by assuming that self-interest and malice, outside of a few exceptional cases, are evenly distributed across tribes, organizations, and political entities, and that when I find a particularly self-interested or malicious person that's evidence about their own personality rather than about tribal characteristics. This is almost certainly false and indeed requires not only bad priors but bad Bayesian inference, but I haven't yet found a way to use all but the narrowest and most obvious negative-valence concepts to predict group behavior without inviting more bias than I'd be preventing.