I definitely don't think the method of picking a single reference class is a good one. Many overlapping big and small classes aren't a problem if you reason correctly.

Re: what if conviction rates are drastically unfair? That's a good point. It's definitely possible to prove that some people who were convicted are likely innocent (but not all of them, just those with e.g. exonerating DNA evidence). I suppose I assumed that the false-guilty rate in the US is less than 10%, but knowing that the false guilty for some privileged class (let's say attractive white folk) is significantly lower, would mean that "X% likely to be guilty when tried" could only come from "X% found guilty when tried" after adjusting as much as possible for the difference in false-guilties and false-innocents.

I agree with you that it's difficult to reason correctly. I'll still encourage people to do so.