It depends on your objective function, doesn't it? Computer models definitely do better than humans on the ostensible objective function --- predicting recidivi

It depends on your objective function, doesn't it? Computer models definitely do better than humans on the ostensible objective function --- predicting recidivism --- but computer models don't have decades of exposure to subtle social cues that signal to humans that the real objective function is one subtly different from the one that a naive understanding of the system would suggest. That's really the beautiful thing about our shift toward algorithms: we'll no longer be able to rely on these subtle social pressures. If we want to optimize a specific function, we need to put that function in the open, where everyone can see it. Then, as the linked article demonstrates, we can compare the idealized and actually-desired functions and perform a real cost-benefit analysis.