A cynic might think that the value of a complicated algorithm is you can, as Pinboard said in a similar context, money launder your bias.
Hell, I do ml on a daily basis, and for anything but a regression or a decision tree, it's complex and/or impossible to even explain why an algorithm picked what it did for a specific example. Let alone to evaluate what an algorithm is picking up on in general.
And that doesn't even get into the mess that is highly correlated variables (eg ethnicity, SES, income, peer income, parental income, education, arrest rate, housing location) most of which are largely synonyms for each other. And the bias in the data themselves -- being poor or minority increases the detection rate of criminality.
