When you think of “civil rights”, many things probably come to mind. Generations-long struggles for racial or gender equality and equal protection for other disenfranchised communities are most commonly associated with the phrase. What you do not think of, most likely, are algorithms. In fact, unless they are of special interest or part of your job, you probably don’t think much about algorithms at all. In our current digital age, however, this may be a mistake. Algorithms have a significant impact on our lives, in spheres ranging from work to recreation, including what we are shown on our favorite websites (influencing how we view the world). Baised algorithms are also creating new forms of discrimination, impacting the very groups that have previously fought so tirelessly to secure equal rights. To understand how this is happening, we explore what algorithms are, how they can perpetuate biases, and the fight being waged to ensure that this technology does not become a new source of inequality. 

What are Algorithms?

Before delving into the impact algorithms have on social justice, it is important to understand what they actually are. The most basic definition is that algorithms are “a plan, or step-by-step instructions on how to solve a problem”. They are extremely important to computer science, and act as one of the four “cornerstones” of computational thinking. Basically, algorithms allow  computers to transform vast amounts of data into useful information, through a process that involves input, computation, and finally, output. Advancements in Machine Learning (ML) have allowed algorithms to learn from past versions and become more complex (and useful). Taken together, algorithms are instrumental to most of the technology that we use today, both on the personal and societal levels. 

Algorithmic Bias

It may seem that algorithms are nothing more than a simple (yet vital) part of the digital ecosystem, helping make our computers work better for us. So how could something so benign be perpetuating biases or even actively harming marginalized groups? The answer actually lies in their ubiquity, as institutions across society increasingly turn to algorithms to streamline operations and advance capabilities. Recent years have brought to light multiple cases of “algorithmic bias”, leading to mostly unintentional (though still harmful) discrimination often caused by the data on which algorithms are “trained” on. Real-world examples of the impacts of this type of bias abound. Amazon, one of the world’s largest companies, came under fire for an AI-based tool developed to sort through resumes after it was shown to be biased against women. Upon further examination, it was determined that the bias was caused by AI “learning” from the resume data of mostly male applicants (reflecting the gender skew in the tech industry) and subsequently giving males preference to advance to the interview stage. Racial bias has also been shown to be perpetuated by algorithms, manifesting in the form of lower credit scores for Black and other minority borrowers, which can result in higher interest rates or denial of loans. In the medical field, AI-based healthcare scheduling software concluded that Black patients were more likely to “no-show” appointments and therefore relegated them to less desirable time slots, leading to 30% longer wait times compared to non-Black patients. In more serious instances, it has been suggested that ML algorithms may be less likely to recommend life-saving treatment options to Black patients and have more difficulty identifying dermatological conditions on those with darker skin tones. As ML advances, these issues will likely continue to worsen. In fact, Data-Pop Alliance’s own recently published research found that ML algorithms applied to earth observation data had a high likelihood of perpetuating gender bias. With equity in employment, financial inclusion, and healthcare at stake, the need to address these biases is clear, leaving only the question of how best to do so.

Algorithmic Justice 

The first steps in fighting against algorithmic bias involve recognizing that, unlike other forms of discrimination, it is often unintentional, and also that just because data is involved (rather than humans) does not mean neutrality can be assumed. From there, concrete steps can be taken to ensure more equitable design. Recognizing that the datasets from which AL/ML “train” are often biased is vital to finding a solution. Sometimes called “dirty data”, these biased samples increase the likelihood of biased algorithms, and should be avoided or compensated for. Luckily researchers are working towards both designing tests to determine if bias exists and, if found, developing methods to compensate for it. Algorithmic transparency is another major tool towards achieving fairness. Currently the “black box” nature of algorithms makes investigation into their inner workings impossible, leading to calls for open access from affected individuals and even government regulators. Government action, such as the UK’s new algorithmic transparency framework, will likely play a major role going forward, but like all struggles for equality, a diverse coalition of stakeholders is needed to meet the challenge. Organizations including the Algorithmic Justice League, The Center for Applied Artificial Intelligence, Data & Society, and many others are taking up the mantle and working to make algorithms more transparent, fair, and inclusive. As awareness grows, there is still hope to address these issues now, in the early stages of Fourth Industrial Revolution, before they have the chance to become ingrained and systemic. 

Join us in the edition of Links We Like as we explore the topic of algorithmic justice and what the future holds in the fight against digital inequality.