Alexandria Ocasio-Cortez Is Absolutely Right About Racist Algorithms

Interesting talk relating to the topic: https://www.youtube.com/watch?v=jIXIuYdnyyk

Many approaches in fair machine learning that try to 'de-bias' the algorithm basically just do stuff like reducing the accuracy in the advantaged group to make the algorithm seem more fair - that is hardly what you want and will just make you susceptible to charges of discriminating against the majority or employing affirmative action. Probably rightfully so, because that's what you do. It's absolutely fine if that's the intent, but then you should have a public discussion where you are open about the fact that you manually tinkered with the parameters to prefer fairness over accuracy (which can probably be a valid goal).

I think finding the problems with the data is very important though. Everyone wins if the quality of your data increases: the algorithm can become both more accurate and also more fair. And it can also identify societal causes for this biased data, for instance police being more sensitive to crimes of minorities, which will then feed back to the innocent algorithm.

A related point is of course that we should be wary of putting too much power and trust into faceless algorithms in the first place.

Also some interesting collection of papers on the matter: https://fairmlclass.github.io/

Maybe you would be interested in something like term-limits for Congress?

If you re-read my comment, it's clear that I'm not talking about rates of representation. You are trying to bring that in. We're talking about the process, be it a policy implementation or an algorithm.

It's the age-old "equality of outcome" vs. "equality of opportunity."

https://archive.org/stream/HarrisonBergeron/Harrison%20Berge...

Computers will tell you you can't get a loan because the algorithm says you won't repay.

Or you can't get this job, or that opportunity, or should be spied on, or whatever.

These may be due to biases in the data (imagine a neural net confusing being black with being poor based on historical data, etc).

Now, I am pretty convinced that this stuff is going to screw us all over -- if you thought human bureaucracy was bad, wait until it's all in the computer and no one has permission or ability to change it.

It seems to me that we may have a rare opportunity here: I'm not sure if this wave of tolerance and caring about minorities is permanent or not, but while people do care, and while computers are provably screwing all kinds of people over, we may be able to get enough people to care so that we can curb some of this stuff. i.e. perhaps don't allow neural nets to predict who is going to commit a crime or who shouldn't have a job, because they're too fallible and too easy to rig.

For example, you can keep feeding it new datasets until it produces the biased result you want and now that's your model, and then you can blame "the algorithm" and "bad data" and not yourself. e.g. maybe you feed it old census data instead of recent data, or old crime data, etc.

Of course, knee-jerk "ban the algorithms" isn't a solution, but starting to talk about and think about a computer bureaucracy where no one accessible to you has permission or ability to change the output of the computer, no matter how exceptional the case may be.

As a software person, I have absolutely no issue with telling other software people to do their job better, and I see no shortage of people that aren't even trying on that front.

Its your comparison between a process that takes safety, reliability and a reduction of harm as a core, first value, and a process that cares first about money, and second about novelty and somewhere way down the line people, maybe, that is offensive frankly.