Computer Scientists Find Bias in Algorithms

First, some definitions:

Equality of opportunity: everyone takes the same test, outcome is dependent on individual, even if the group containning that individual scores differently than other groups

Equality of outcomes: everyone takes the same test, but the individuals scores are weighted based on the group of which they are a member, so that representation of groups is equivalent across all groups.

Further, we can separate bias in terms of intentional and unintentional, where the former indicates direct manipulation of selection criteria in order to change group outcomes to match the manipulator's desires. The latter occurs when systematic bias affects the individual's scores, such as when a particular group has less access to educational opportunities.

The article values equality of outcomes, by encouraging algorithms to take into account the individual's group when performing selection. But in cases of unintentional bias, this can mask problems. In the long term, it would be more beneficial to ensure groups have equal access to education than to upwardly adjust their scores.

In fact, this upward bias to adjust for perceived unintentional lowered outcomes can in fact perpetuate the initial biases: the message is that a particular group underperforms, and thus needs extra help.

This is totally ridiculous. All it does is remove the useful data from the dataset and makes it's predictions worse.

Imagine are using an algorithm to predict how likely someone is to pay back a loan, or make an insurance claim, or perform as an employee, etc. Things like their grades or education would correlate with that. but it might also correlate with race or gender, so the algorithm would be labelled "racist" and "sexist".

Even though it has nothing to do with racism whatsoever. A white person and a black person with the same education would get exactly the same result. The algorithm is totally fair and meritorious.

The true solution would be to condition on the race or sex. So the algorithm predicts the probability they will be paid back given that they are male or female. Then in production you average the results of each possible sex and race together.

So if women were really better drivers than men, they would still pay the same rate. The algorithm would know that they are better drivers, but it would treat everyone gender neutral. But it would still take into account education or other variables.

Although it's not clear to me why you would want to do this. If men really are more risky drivers than women, they should pay more. And they do. If they really are less likely to pay back a loan, then they should have a higher interest rate.

In any case, humans make decisions about stuff like this all the time and they are way, way, worse and more biased. In most cases very simple statistics vastly outperform the best humans: http://lesswrong.com/lw/3gv/statistical_prediction_rules_out...

"If the test results show that the algorithm under question can distinguish attributes—like whether a data point represents a male or female—then it’s labeled biased. The other algorithm the researchers introduced tries to remedy the bias by modifying the actual data set so that any selection algorithm would deliver fair results. The algorithm does this by blurring attributes that may be correlated to, say, race or gender."

So if you're testing construction workers, you identify bias, then you select strength as the variable causing bias, and you adjust the strength values of the dataset or force the algorithm to devalue strength?

Racism/sexism are real and we need new ways to fight them, but this is like a right wing satire of affirmative action.

The basis of the argument is that if decisions correlate with race, than the cause of the decision is race. This is of course wrong. Correlation does not imply causation and the algorithms cannot even have a concept of race or sex unless that data is deliberately and explicitly placed into the data set.

A realistic complaint would be that these algorithms dehumanize people and create a weird algorithm based bureaucracy that mimics The Trial.

Maybe someone is trying to conflate the two arguments so they can dismiss the second one by pretending it's the first?