it should go without saying that complicated decision-making algorithms need human supervision. but this WaPo article making the rounds completely undermines its own argument by picking an example that is in almost every sense a failure of human "experts" https://www.washingtonpost.com/technology/2020/12/23/covid-vaccine-algorithm-failure/

if we consider humans rationally formulating bad decision rules to be an instance of algorithmic failure, then the only negative outcomes that can't be blamed on algorithms are those produced by opaque holistic human judgment. i don't think we want to encourage more of that

concerns about the growing role of AI in day-to-day decision-making deserve to be taken seriously. but the problem is *not* that humans should be making more of the decisions; it's that humans should be thinking more carefully about how to evaluate and regulate algorithms

put differently, the question "should we trust algorithms?" carries the soft implication that we were already doing okay trusting human experts. we weren't. human decision-makers regularly make terrible, hugely biased decisions. we should trust neither humans nor algorithms

if anything, the benefit of algorithmic decisions is they can (at least in principle) be interrogated in a way that human decisions can't. you will never know for sure why your Uber driver suddenly swerved and crashed; but Tesla *might* be able to explain an Autopilot crash

one irony of the backlash against algorithmic decision-making is that many of the biases people object to were already present—often to a far greater extent—in human decisions, and this didn't seem to bother nearly as many people

is it bad if a racist algorithm gives some groups disproportionately long jail sentences? of course it's bad. but it's also bad when racist human judges do the same—and somehow there hasn't been the same movement to evaluate and regulate human decision-making in this context

to be clear, i'm not saying "if you didn't complain about racist judges, you shouldn't complain about racist algorithms"; you should complain about both! the point is that the root problem is not the use of algorithms, it's the lack of appropriate testing/regulation regimes.