FOUR things to know about race and gender bias in algorithms:
1. The bias starts in the data
2. The algorithms don't create the bias but they do transmit it
3. There are a huge number of other biases. Race and gender bias are just the most obvious
4. It's fixable! π§΅π
1. The bias starts in the data
2. The algorithms don't create the bias but they do transmit it
3. There are a huge number of other biases. Race and gender bias are just the most obvious
4. It's fixable! π§΅π
By race and gender bias in algorithms, I mean the tendency for heavily data-driven AI algorithms to do things like reproduce negative stereotypes about women and people of color and to center white male subjects as normal or baseline. 2/9 
While race and gender bias in algorithms *is fixable*, the current fixes aren't easy. They require us to understand and then mathematically model the processes that generate the biases in the data in the first place. 3/9
Traditionally, statistics culture says any bias in the model is the responsibility of the modeler. Modelers are expected to address potential bias to the best of their ability. This isn't race or gender bias specific. Statisticians care a lot about bias of all kinds. 4/9
There doesn't seem to be a consensus in the computer science and machine learning communities regarding who is responsible for misbehaving algorithms, although many AI ethics researchers are pushing for these communities to take more responsibility. 5/9
Many companies make massive profits from mathematical models which contain gender and race bias so one could argue that they have some ethical responsibility to make good faith attempts to address these biases. 6/9
There are many techniques in statistics for addressing bias in data but they require you to understand the data on a deeper level than is common in machine learning applications so they are hard for machine learning researchers to adopt. 7/9
Machine learning focuses a lot on the parts of statistics that can be automated and it's hard to automate a deep sociological understanding of race and gender and the cultural processes that cause bias in the production data. 8/9
While fixing bias in machine learning algorithms is a solvable problem, the cost and complexity of fixing it will probably mean that such biases will remain an issue for some time. 9/9
I hope you found this thread informative. If you would like to support this kind of content, follow me and also like and retweet the thread. π
Some caveats on "algorithms don't create the bias":
1. I'm assuming people aren't biased in how they made the algorithm. I think this type of bias falls into the typical structural biases we see in society and isn't tech specific so I didn't address it in this thread.
1. I'm assuming people aren't biased in how they made the algorithm. I think this type of bias falls into the typical structural biases we see in society and isn't tech specific so I didn't address it in this thread.
2. Algorithms can technically create bias by just being crappy algorithms or due to mistaken assumptions about the data. I didn't address this situation in my thread either since making mistakes in data analysis isn't tech specific either.
One thing I want to make clear is I don't think fixing the biases in the data will address all the issues with gender and race discrimination in AI algorithms. Some people naively think this and it's almost certainly wrong.
Oof. This thread is doing numbers. π There are a few comments disagreeing with this thread. One thing to keep in mind as you read them. As far as I can tell, they are misinterpreting what I said because theyβre using a different definition of bias.
They see an algorithm as βbiasedβ if it has tendency to create unequal outcomes. Iβm using bias in the statistical sense. Basically, Iβm talking about how to create a colorblind algorithm that doesnβt use race in its decision making.
I think their concerns about fairness are valid but itβs not at all what I was talking about.
β’ β’ β’
Missing some Tweet in this thread? You can try to
force a refresh
γ
