[makomk]

No, they show that a completely inhuman intelligence, designed to learn to show humans what they want to see, can successfully cater to human bias. For example, the Salon article talks about Google AdSense showing ads for arrest records when someone searches for black-sounding names but not white-sounding ones. Google are quite open about the fact that they choose the ads they think are most likely to be clicked. So if people are more likely to click on ads for arrest records when looking for information on a black person, that completely inhuman intelligence, with no intrinsic biases, will happily cater to their racism. That you argue this somehow means racism is a useful, predictive heuristic of anything other than how racists act says a lot.

[yummyfajitas]

Another example from the article shows that men are more likely than women to click on an ad for a high income job, or that low income people are going to click on ads for high interest loans. These are stereotypes that seem to be confirmed by the algorithm.

The core question - do you believe the problem will be fixed by better machine learning algorithms? Going back to the current example, do you believe that a Bayes-optimal machine learning algorithm for predicting criminal behavior will be "unbiased" (in the sense of social justice, not in the sense of statistics)?

Or, more concretely, do you believe that the only problem that mtgx and smtddr are complaining about is that our ML algorithms aren't good enough and that maybe we need deep learning instead of random forests?