Bias and discrimination hidden in algorithms
Introduction
Digital technologies are gradually penetrating into people’s lives. For example, thousands of users are active on the various digital platforms to search, collect and deliver information every day. According to Statista, In the fourth quarter of 2021, Facebook had 2.91 billion monthly active users, and the number of Google users worldwide is nearly four billion. These are huge numbers, which prove that digital platforms are very attractive to people. In addition, the increasing number of individuals, organizations and governments rely on artificial intelligence equipment to gain convenience and save time. In daily life, users rely on facial recognition technology to conduct transactions, like Alipay. As well as, some organizations, such as Police stations, depend on facial recognition technology to track specific people. Hence, digital technologies have become an important part of human activity.
Whether it’s a digital platform or an artificial intelligence device, there are sophisticated and complex algorithmic systems behind these emerging digital technologies to manipulate. The algorithm performs calculations and data processing specifications through a specific set of formulas. It includes the processing and analysis of digital data, and digitization and archiving projects that aimed at transforming analog texts and physical objects into organizable and searchable digital forms and performing aggregate, automated, or guided analyses in their basic form (Schnapp and Presner, 2009, as cited in Sadowski, 2019). Undeniably, algorithmic systems help and facilitate human activities by mining, collecting, analyzing, processing, and executing data in several complex processes. When users adopt an algorithmic decision, their cognition and behaviors are also potentially influenced by the algorithmic decision. For instance, When users are unfamiliar with a particular area, they often use some search engines, such as Google or Bing, to get information and answers. These search engines will lead the user to a specific interface. Through these specific interfaces, the user establishes knowledge networks about this specific area. If the search engine is not neutral, the user will also have a certain biased impression of this unfamiliar field. According to Jackson(2018), algorithms that are hidden from view are quantifying our lives. Hence, those emerging technologies driven by algorithmic systems are gradually shaping some cognition and behavior of individuals. This blog will discuss algorithms that may contain some biases and discrimination that can have a significant impact on people’s lives.
Algorithmic Bias
The results deduced by the algorithm are not neutral, and even the setting of the algorithmic system is not entirely neutral. This is because data is never raw. In scientific research, choosing what to measure and how to measure is fundamental. But in many cases, especially in the social sciences, there is no clear measure of what we want to capture. We have to “operate” in a certain way, which means that we have to standardize some techniques for measuring it (Nick, 2018). Therefore, when collecting data, some things may be missed, and some things may be redefined. This lead to some deviations in the process of dealing with data. This also affects the outcome of algorithmic decisions. In addition, data mining is based on algorithmic models that repurpose data to create new meaning or infer meaning from incomplete data (Jackson, 2018). Because algorithmic systems are designed by humans, it are influenced by various factors, such as political, economic, cultural, social, etc. According to Bozdag(2013), humans not only shape algorithm design, but they may also manually influence the filtering process while the algorithm is running. The process of excavating and collecting data is limited and the design of algorithmic systems is also not objective. These allow for some bias and discrimination to be unconsciously embedded in the ideology of the algorithm. Algorithmic models can learn patterns that correspond with social characteristics (e.g., race or gender) and yield biased results, whether intended or not, by exploiting historical data and inferring meaning from absent data (Greenwald, 2017; Williams et al., 2018, as cited in Jackson, 2018). It has been demonstrated by many scholars that algorithmic systems are biased. Bias can appear in a computer system in a variety of ways: pre-existing bias in society can influence system design, technical bias can originate due to technical constraints, and emergent bias can appear after software implementation is done and deployed (Friedman and Nissenbaum 1996, as cited in Bozdag, 2013). Several writers have demonstrated how search engines, particularly in terms of coverage, indexing, and ranking, can contain technical biases (Van Couvering 2007; Diaz 2008; Mowshowitz and Kawaguchi 2002; Vaughan and Thelwall 2004; Witten 2007, as cited in Bozdag, 2013).
The bias is hidden in Google’s algorithm
It is acknowledged that Google is a platform with powerful search capabilities and recommendations. However, it is also a classic example to show the prejudice and discrimination designed in algorithms to increase racism and sexism (Hoffman, 2018). For instance, in the past, Google searches have provided pornographic pages related to particular racial groups or ethnicity, especially for black women. In addition, Google provided advertising and search results with criminal backgrounds that appeared to be based on African-American names, including ads that targeted women whose identities have been maligned in the media, such as black women and girls (Hoffman, 2018). Furthermore, an MIT researcher found that people with “black-sounding” names (DeShawn, Darnell, Jermaine) were more likely to have arrest records than those with “white-sounding” names (Geoffrey, Jill, Emma, Sweeney, 2013, as cited in Jackson, 2018). Moreover, Google shows different types of advertisements according to gender. When users want to find some advertisement about finding a job, men often were shown advertisements that encouraged them to seek coaches for higher-paying jobs, while women were shown generic and normal advertisements (Datta, Tschantz, Datta, 2015, as cited in Jackson, 2018). Using Google Translate also show sex discrimination. For example, Turkish language, “o bir doktor” was interpreted by Google Translate as “he is a doctor,” and “o bir hemşire” was interpreted as “she is a nurse,” representing the role of the male as a doctor and the role of the female as a nurse (Bomstein, 2017, as cited in Jackson, 2018). This means that in Google’s algorithmic ideology, “male” associate with the character of a doctor and “female” with the character of a nurse. Google as “gatekeepers” of cyberspace, could direct hundreds of millions of users to particular content, but not to other content, and guide into certain sources, but not to other sources. According to Bozdag(2013), What a daily user sees and do not see is influenced by search and recommendation engines. Hence, users may consciously or unconsciously acknowledge those informations with bias. This may enhance social bias and discrimination.
Facial Recognition Harm
Facial recognition is an emerging digital technology that to recognize a human face. To identify a match, facial recognition compares the information to a database of known faces. It is possible to identify people in photos, videos, or in real-time using facial recognition systems. Facial recognition can aid in the verification of a person’s identification. However, not all faces are recognized by facial recognition systems. Data sets used to train facial recognition technologies are not diverse. Some facial characteristics are ignored, and some facial characteristics are redefined. This may create a feedback loop of further discrimination.
In 2015, Google’s algorithmic photo recognition software labeled a photo of a black developer, Alciné, and his friend in New York as a gorilla. Yonatan Zunger, an engineer and chief architect at Google+, was quick to respond to the incident on Twitter saying he was very sorry and promising that Google’s Photos team was working on a fix and that it would not happen again. However, in the process of fixing the program, it was found to be ineffective. Zunger said that Google was working on “long-term fixes”, including “better recognition of dark-skinned faces”. Later, Google spokeswoman Katie Watson issued another statement saying she was shocked and truly sorry that this had happened. Alciné response that he understand how this happened, but the core of the question is “why”.
Many commercial facial recognition systems have difficulty or misrecognition when detecting darker faces. This reflects an element of digital epidermalisation whiteness privilege, on the other word, white faces are the standard for the system (Buolamwini and Gebru, as cited in Stark, 2019). The reasons for causing those issues are related to the group of engineers. According to Stark(2019), Facial recognition technologies were neither invented for a social vacuum, they were designed and operated by real people to classify real people. This means that if the group of engineers designing the facial recognition algorithm is homogeneous, then the results of facial recognition is not diverse. For example, several major technology companies, including Apple, Twitter, and Slack, have released diversity statistics showing that their workforce is overwhelmingly white and male, especially when it comes to technology jobs (Ferro, 2016). For example, there are 72% of white and 82 % of males in engineer department in Slack. This means that women of colour make up 9 % of Slack’s engineers, or half of the company’s total female engineers (Ferro, 2016). The data captured by the algorithmic system lack diversity, and this situation must also influence the algorithmic decisions. Owing to the group of algorithmic designers, facial recognition system is difficult to classify faces that do not look like the training set’s standard (white) face.
Conclusion
In this blog we discuss whether bias and discrimination exist in algorithms. Algorithm systems are inherently biased. Firstly, dealing with data is a complex process and different ways of processing data can lead to different results. The process of collecting data is not visible and the data are handled in a special way (Hoffman, 2018). Those data are fed into an algorithm to calculate and deduce. Small deviations can affect algorithmic decisions. Secondly, the group of engineers that create algorithms is not diverse. Their biases, attitudes, and assumptions, which are overwhelmingly white and male, tend to shape algorithm design (Karppinen & OinasKukkonen, 2013; Rayóme, 2018; Yetim, 2011, as cited in Jackson, 2018). This results in algorithmic systems and algorithmic decisions that are not completely neutral and objective.
We enter in the tremendously creative era, with the majority of human activity shifting from analog to digital. Digital technologies are shaped by humans and in turn affects human life. Whether it is influential digital platforms such as Google and Facebook, or some AI technologies that can bring huge convenience to human life, like face recognition, they all have a huge impact on users. Algorithmic decisions may affect users’ awareness and judgment. Algorithmic bias may harm the interests of some marginalized groups. As Jackson(2018) mentions that although biases influence everyone, the degree of impact is different. biases impose a greater burden on members of marginalized communities (e.g., racial and ethnic groups). Therefore, the stereotype that data is neutral or algorithms are neutral should be broken. The way in which the digital world is constructed may have long-term implications for human activity. Identifying and resolving biased algorithmic results will also be a serious challenge in the future.
References
Bozdag, E. (2013). Bias in algorithmic filtering and personalization. Ethics and Information Technology, 15(3), 209–227. https://doi.org/10.1007/s10676-013-9321-6
cottonbro. (2021a, May 27). Retrieved April 7, 2022, from Pexels website: https://www.pexels.com/zh-cn/photo/8090124/
Facebook MAU worldwide 2021. (n.d.-b). Retrieved April 8, 2022, from Statista website: https://www.statista.com/statistics/264810/number-of-monthly-active-facebook-users-worldwide/
Ferro, S. (2016a, March 1). Here’s why facial recognition tech can’t figure out black people. HuffPost. Retrieved from https://www.huffpost.com/entry/heres-why-facial-recognition-tech-cant-figure-out-black-people_n_56d5c2b1e4b0bf0dab3371eb
Guynn, J. (2015b, July 1). Google Photos labeled black people “gorillas.” USA TODAY. Retrieved from https://www.usatoday.com/story/tech/2015/07/01/google-apologizes-after-photos-identify-black-people-as-gorillas/29567465/
Hoffman, A .(2018). Data Violence and How Bad Engineering Choices Can Damage Society Data Violence and How Bad Engineering Choices Can Damage Society | by Anna Lauren Hoffmann | Medium
Jackson, J. R. (2018). Algorithmic Bias. Journal of Leadership, Accountability and Ethics, 15(4), 55-65.
Nick Barrowman. (2018). Why Data Is Never Raw. New Atlantis (Washington, D.C.),(56), 129-135.
Stark. (2019). Facial recognition is the plutonium of AI. Crossroads (Association for Computing Machinery), 25(3), 50–55. https://doi.org/10.1145/3313129
Sadowski. (2019). When data is capital: Datafication, accumulation, and extraction. Big Data & Society, 6(1), 205395171882054–. https://doi.org/10.1177/2053951718820549
Noble, S. (2018c, March 26). Google has a striking history of bias against black girls. Time. Retrieved from https://time.com/5209144/google-search-engine-algorithm-bias-racism/