What is Algorithmic Bias?
Algorithmic bias, in the simplest terms, is a systematic error in predictive computation. In some contexts, the term bias describes statistical mistakes that predictive models make because of code bugs, poor model selection, inappropriate optimization metrics, or suppressed data. Some also define algorithmic bias as probabilistic bias, or how likely a person from group A is to get the same outcome as another person in group B.
Most commonly, especially among non-data-scientists, algorithmic bias refers to computational discrimination whereby unfair outcomes privilege one arbitrary group of people over another. In this definition, the focus is on the disparate impact technology may have that reinforces social biases based on race, gender, sexuality, ethnicity, age, and disability. Algorithmic bias is not just deplorable, in many cases it is illegal. Discrimination on the basis of protected categories is prohibited in the US. However, cases against discrimination have the burden of proving an outcome was because of that discrimination. This burden of proof often allows programmers and corporations to shirk responsibility and repercussions.
In reality, algorithmic bias encompasses all of these ideas as they are mathematically and philosophically defined. However, we explicitly ground this guide in a larger intersectional social justice framework. We are arguing that any algorithms that perpetuate the status quo and do not improve social equity for people of color, women, the poor and working class, those with disabilities, and other marginalized groups are biased. Further, we hold that any attempt to do data science in an "apolitical" way is inherently politically conservative, because data science as currently practiced has shown time and time again to harm vulnerable communities. In other words, we center justice in our definition of algorithmic bias. Machine learning and AI should not only work equally well for all people, but all software should also aspire to support a larger struggle for economic, racial, and gender equity.
In many ways, the term "algorithmic bias" is misleading. There is no single algorithm that runs the entire service within any website or app. Instead, these services are a complex network of software programs and data inputs/outputs that determine a single end user's experience. Part of why algorithmic bias is so pernicious is that it is challenging to pinpoint where something could or did go wrong in a project's codebase.
Similarly, algorithmically driven products exist within the complex social fabric we each must navigate. To properly assess algorithmic bias, we must understand the social structures that have empowered a techno-elite class to determine masses of marginalized communities' fates. These systems have led to computers having the power to criminalize, deny access to resources, and just plain inconvenience people due to their skin color, gender, sexuality, age, or ability status. We will never escape the horrors of biased algorithms until our culture adopts a technology development framework—an ethos—that both meaningfully grapples with historical and present-day injustices and prioritizes community empowerment and self-determination.
