Can Bias in Artificial Intelligence be Ethical?
1 Introduction
Science fiction has often associated Artificial Intelligence and related technologies with far-fetched, unbelievable futures that have only been pictured in dystopian novels. However, this view is no longer in the realm of futuristic science fiction as artificial intelligence and machine learning (AI/ML) technologies increasingly become a part of our everyday lives. These smart technologies have not only made their way into the average household in the form of Amazon’s Alexa, Apple’s Siri, etc., but many businesses and enterprises have completely shifted to AI-based infrastructures. AI has enabled companies to process and analyze extraordinary amounts of data, streamlining the decision-making process and significantly increasing efficiency, productivity, and innovation. The business capabilities of AI technology can be broadly encompassed by three key needs: automating business processes, gaining insight through data analysis, and engaging with customers and employees. While AI technology has brought great value to each of these sectors, it has also brought about new problems. Some examples include displacement of jobs due to automation and inaccuracy in decisions caused by limited datasets. One issue which has been overlooked that is critical to the development of AI/ML technology is biased datasets and algorithms leading to biased outcomes — whether such bias is conscious or unconscious. The intricacies of this issue can be seen in its impact on social media and search engine companies.
These companies are and have always been, strong proponents of AI/ML technologies, regularly releasing new products and features on their platforms which utilize these algorithms. These platforms store and analyze prodigious amounts of user data acquired through a comprehensive record of a user’s clicks, likes, shares, and other actions performed on websites and applications. AI/ML algorithms are implemented to process that data and generate content that is more relevant for each individual user. While the intended goal of using these algorithms is to make things like advertisements, search results, and search suggestions more personalized, they do not always succeed. When developing and training AI/ML systems, the training data that is used is inherently biased (virtually all data is), which results in inherently biased systems. If the data or algorithmic bias is not filtered or corrected for, the system can yield results that may be harmful to users, contrary to their intended effect.
With the emergence of a new era centered around AI/ML technologies, algorithmic bias has become a rising concern, particularly when such bias impacts protected classes (defined by aspects such as sex, race, religion, color, national origin, age, people with physical or mental handicaps, familial status, veteran status, and genetic information). The most detrimental class of bias is latent bias, where an algorithm insidiously and unintentionally correlates results with sensitive protected information (SPI) such as gender, race, sexuality, income, etc. As AI/ML systems become more prevalent in the commercial realm, it is critical that latent bias is identified and mitigated by the developers of those systems. Identifying biases pertaining to gender or race is of particular importance, as these are arguably the most prominent examples of harmful biases. From computer vision systems to language processing algorithms, AI and ML systems can often be unintentionally skewed against racial minority groups and women.
While it is important that we combat these harmful algorithmic biases, we also must recognize the difference between detrimental and beneficial biases. For example, if a man is shopping for clothing online and searches for men’s clothing, a search algorithm would likely take this information into account and skew the user’s future search results toward men’s clothing. This would be a beneficial bias as it would hopefully improve the efficiency and personalization of the user’s shopping experience. Most would agree that there is a utility in this bias, which is why AI/ML algorithms are so widely implemented in search algorithms. However, if an algorithm is intended to filter through applicants for a particular university, for example, and that university has a predominantly male population, the algorithm might recognize a relationship between being male and being admitted and subsequently factor gender into admission. In this situation, the algorithm is clearly wrong, but while this outcome is vastly different from that of the previous example, both situations are results of the same fundamental premise that AI/ML algorithms recognize patterns and act in accordance with those patterns. This is what makes algorithmic bias such a dilemma — how do we combat the negative aspects of bias while maintaining or improving the overall utility of AI/ML algorithms? While bias is not inherently unethical, and while developers generally do not intend to introduce negative biases into the algorithms they train, many algorithms in use today are susceptible to unethical biases, and algorithms that are trained using large datasets often display unethical behaviors with regard to racial and gender discrimination. In this paper, we argue that the overall state of AI/ML algorithms in terms of allowance for bias is unethical, as there is still much work to be done with regard to mitigating negative biases and defining what kinds of bias are acceptable. We will present and analyze specific cases of algorithmic bias in regards to both race and gender as well as discuss potential solutions to combat AI-introduced discrimination.
2 Racial Bias
Perhaps one of the most infamous examples of racial bias in AI/ML is the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) recidivism algorithm. According to ProPublica analysts, COMPAS is an algorithm that processes data gathered from a questionnaire filled out by defendants to predict the “likelihood of each committing a future crime.” This prediction, called a “risk assessment,” is provided to judges in a number of states to help inform their decisions during criminal sentencing (Angwin et al. 2016). As the ProPublica article which exposed COMPAS’s bias explains, “The appeal of risk scores is obvious: The United States locks up far more people than any other country, a disproportionate number of them black. For more than two centuries, the key decisions in the legal process, from pretrial release to sentencing to parole, have been in the hands of human beings guided by their instincts and personal biases.” As such, in order for risk assessment algorithms to be useful, it is clearly very important that risk assessment algorithms are as unbiased and accurate as possible. The ProPublica investigation found two key results which shed light on the racial bias hidden under COMPAS’s supposed utility: the algorithm was nearly twice as likely to incorrectly label black defendants as future criminals as white defendants, and was also significantly more likely to mislabel white defendants as low-risk for recidivism than black defendants (45% to 23% and 48% to 28%, respectively). Basically, the algorithm disproportionately targeted black defendants as high-risk for recidivism and white defendants as low-risk.
While these statistics illustrate a jarring issue with COMPAS’s decision-making process, a Washington Post reanalysis of the data ProPublica used shows that the line between a fairly and unfairly biased algorithm is much fuzzier than ProPublica makes it out to be. The article explains that recidivism rates for black and white defendants within each individual risk category are approximately the same, and that “The overall recidivism rate for black defendants is higher than for white defendants (52 percent vs. 39 percent).” It then makes the argument that the unfairness exposed by ProPublica is “mathematically guaranteed” as a result of the aforementioned facts about recidivism rates. The explanation is that “If the recidivism rate for white and black defendants is the same within each risk category, and if black defendants have a higher overall recidivism rate, then a greater share of black defendants will be classified as high risk [and, therefore,] a greater share of black defendants who do not re-offend will also be classified as high risk” (Corbett-Davies et al. 2016). This, of course, does not deny the fact that by ProPublica’s standards, which are arguably reasonable, COMPAS is a racially biased algorithm that unfairly assesses defendants with regard to race. However, the fact that recidivism rates for black and white defendants is the same in each risk category means that, from another perspective, COMPAS is an algorithm that does indeed treat defendants fairly across racial lines. This makes the issue of determining the ethicality of COMPAS’s algorithm extremely difficult, as there is no objective frame of reference for how the algorithm should be scrutinized in terms of fairness. From ProPublica’s perspective, the formula is highly flawed and disproportionately harms black defendants, but from Northpointe’s perspective, the algorithm is just as unfair for black defendants as it is for white defendants.
This concept is the crux of our thesis: how can we evaluate bias in AI/ML systems if we do not have a standard evaluation process? In the agriculture industry, for example, organizations exist to label certain foods as organic or as non-GMO, and these organizations go through a systematic process with each product they evaluate — they have a standard process to determine whether or not to label a product a certain way. In order for developers to successfully mitigate harmful algorithmic biases, they, too, should have a particular standard or metric to evaluate what kinds of biases are acceptable and to what degree. Currently, there do exist some government organizations such as the National Institute of Standards and Technology who are actively researching to determine ways of assigning a metric on bias, but the wheels of the government turn slowly, meaning there is no immediate solution available yet.
Another example of racial bias in AI/ML is an algorithm “widely used in US hospitals to allocate health care to patients” (Ledford 2019). In a 2019 study led by Ziad Obermeyer, researchers discovered “evidence of racial bias in [the] algorithm, such that Black patients assigned the same level of risk by the algorithm are sicker than White patients,” meaning that millions of black Americans could potentially have been overlooked when they actually required specialized care. According to the researchers, the algorithms based their predictions on how much individuals spend on healthcare annually, which seems like a sound measure of medical need because people with more or worse conditions will presumably spend more on healthcare than those with less or milder conditions. This assumes that a significant amount of the variation in healthcare spending can be directly accounted for by healthcare needs, but in reality spending can vary as a result of many different factors — one of which is race. Race-related factors which may affect healthcare spending include, for example, the fact that white patients are more likely to be able to afford healthcare than black patients, or that black individuals may be more likely to “distrust… the health-care system” (Ledford 2019). Thus, while it is tempting to say that the algorithm itself is racially biased, in reality the data used to train the algorithm was racially biased, so the algorithm “falsely concludes that Black patients are healthier than equally sick White patients” (Obermeyer et al. 2019).
After the release of this study, the company which developed the algorithm worked with Obermeyer’s team “to find variables other than healthcare costs that could be used to calculate a person’s medical needs, and repeated their analysis after tweaking the algorithm accordingly. They found that making these changes reduced bias by 84%” (Ledford 2019). It is important to note that this extraordinary improvement was not a result of modifying the algorithm itself, but rather of changing which variables the algorithm utilized to make its predictions. All of the data that could be fed into the algorithm has some level of bias, but Obermeyer’s team was able to find and utilize data with less of a relationship to race, resulting in less biased predictions from the algorithm.
This leads to another interesting and significant observation: neither COMPAS nor the healthcare algorithm explicitly factored race into their predictions. The COMPAS questionnaire did not ask defendants about their race, and while the healthcare algorithm was fed spending data that has a relationship with race, it also did not specifically use race to determine its recommendations. Thus, it is clear that the subject of bias, which in this case is race, in an AI/ML algorithm does not need to be introduced to that algorithm explicitly for the bias to exist. Simply put, an algorithm does not need to know an individual’s race to be biased toward or against that individual as a result of his or her race. Since racial discrimination is so deeply engrained in our society and has been for a long time, race, unfortunately, is correlated with a plethora of other variables one might use in conjunction with an AI/ML algorithm, like income or health, for example. This means algorithms are extremely susceptible to racial bias if developers are not careful when deciding what data their algorithms should analyze, as exemplified by the healthcare algorithm’s bias and the astronomical reduction in bias achieved by using different variables than healthcare spending.
Addressing these issues is especially difficult because algorithms and data are often considered proprietary by their developers. As a result, if an algorithm is determined by researchers to have racially biased performance, those researchers typically do not have access to the algorithm and data in question and, although they can pinpoint and study the performance issue, they cannot determine why the algorithm behaves the way it does in the first place. Algorithmic transparency, defined by Nicholas Diakopoulos and Michael Koliska as “the disclosure of information about algorithms to enable monitoring, checking, criticism, or intervention by interested parties,” has always been a controversial issue in the fields of artificial intelligence and machine learning. However, similarly to the aforementioned metric for acceptable levels/kinds of bias, there is no consensus as to how transparent organizations must be with regard to their data and algorithms, allowing companies full control over whether or not to address biased systems. For the healthcare algorithm case, Obermeyer and his team were fortunate enough to be given access to “reams of sensitive health data” in order to draw their conclusions and subsequently help improve the algorithm’s performance (Ledford 2019). However, the ProPublica investigators involved in discovering COMPAS’s racial bias were not so lucky, as “Northpointe has refused to disclose the details of its proprietary algorithm, making it impossible to fully assess the extent to which it may be unfair” (Corbett-Davies et al. 2016). In addition to the lack of a standard metric for AI/ML bias, the absence of definitive guidelines for data and algorithmic transparency hinders progress toward reducing algorithmic bias.
A third example of racially biased algorithms is the Google Photos photo-labeling system. Google Photos has a feature that allows users to search their photo gallery using labels that Google’s algorithm applies to each photo or video. According to a 2018 story in Wired Magazine discussing the algorithm’s racial bias, “In 2015, A black software developer embarrassed Google by tweeting that the company’s Photos service had labeled photos of him with a black friend as ‘gorillas.’” This is obviously a terrible and hurtful fault in the algorithm which had to be addressed quickly, so Google responded by “erasing gorillas, and some other primates, from the service’s lexicon” as a quick fix (Simonite 2018). However, the company still has not figured out how to solve the problem, and in fact, the Wired article states that “A Google spokesperson confirmed that ‘gorilla’ was censored from searches and image tags after the 2015 incident, and that ‘chimp,’ ‘chimpanzee,’ and ‘monkey’ are also blocked today.” This highlights the unfortunate but unavoidable reality that while AI/ML has come a long way in recent years, algorithms “can’t easily go beyond the experience of [their] training… [and] even the very best algorithms lack the ability to use common sense, or abstract concepts, to refine their interpretation of the world as humans do” (Simonite 2018).
3 Gender Bias
The increased presence of AI/ML applications in people’s everyday lives has had great influence over their actions, thoughts, and behaviors. However, since the gender of the developers of AI/ML technologies is predominantly male and datasets used to train AI/ML engines have had historically inherent male-skewed gender bias, this disparity is reflected in the output of these engines. Unlike machines, humans have developed the understanding to make decisions based on a combination of factual data and personal experiences, whereas machine decision-making abilities are primarily reliant on identifying patterns from large datasets. If those datasets mirror stereotypical ideologies of gender, the AI/ML machine will unconsciously perpetuate this latent bias.
The biggest assumption that companies make when releasing AI tools and products is that they are inherently neutral. But, “automated systems are not inherently neutral. They reflect the priorities, preferences, and prejudices — the coded gaze — of those who have the power to mold artificial intelligence” (MIT Gender Shades). This truth has been observed more and more in recent years as companies and organizations continue to incorporate AI into many work processes, especially in recruiting and talent management. With hiring algorithms in particular, the severity of the negative implications of gender bias in AI/ML algorithms has been revealed. A prime example is Amazon’s experiment with a new AI-driven recruiting engine that was intended to speed up the identification process for top talent. The goal was to develop an engine that could analyze hundreds of applicants in a matter of minutes and choose the top five best applicants for the position. The computer models that were used to develop this tool were trained using past resumes submitted by job applicants over a ten year period. This methodology, however, did not account for the historical patterns of male-dominance across tech jobs and further reinforced this pattern while choosing candidates (Dastin 2018). In addition, the engine began to penalize resumes that mentioned female organizations or all women’s colleges. The most interesting correlation that engine made, however, was with the language that was used in the resumes. The models driving the engines identified certain words and terminologies to be more masculine than others and favored candidates that utilized such language. This not only resulted in an overall male-dominated hiring pool but also recommended under-qualified candidates since language is not an indicator of technical skill (up until this point most research studies have focused on remedying bias without consulting years of studies on gender ideology present in language).
As companies like Goldman Sachs, Microsoft, IBM, etc. continue to push for the use of AI technology in hiring processes, it is critical to ensure that AI/ML algorithms in use address the evident inequity of opportunities available for women compared to their male counterparts. In the Amazon situation, since gender bias was evident in the data used to train the algorithm, it learned to be biased toward a particular gender even though the algorithm is meant to be more objective than humans. The observation about bias intrinsic in our language illustrates how widespread human biases are, and therefore how easy it is for seemingly objective data to teach bias to AI/ML systems.
Researchers Amit Datta, Anupam Datta, and Michael Carl Tschantz from Carnegie Mellon University and the International Computer Science Institute performed another study pertaining to gender bias in 2014, analyzing Google’s targeted advertising algorithm on third-party sites. Their study focused on how Google’s Ad Settings, an interface which “displays inferences Google has made about a user’s demographics and interests based on his browsing behavior” and allows users to interactively modify or delete those interests in order to change their advertising experience (Datta et al. 2014). Datta’s team focused specifically on the transparency of Google Ad Settings and the level of choice it provides to users using a tool they created called AdFisher, which simulates new users by creating personas from fresh browser instances with no browsing history, cookies, or other personalization. In addition to creating personas, AdFisher also collects and categorizes advertisements in Google’s network displayed to the simulated users. Datta’s team conducted an experiment in which they converted a number of male personas to female using a setting available on Google’s Ad Settings page. Once the switch from male to female was made, there were fewer instances of high-paying job advertisements displayed to the user (Datta et al. 2014). Although it is clear that the system carries some biases against women, as listing the user’s gender as female resulted in lost opportunities compared to listing the user’s gender as male, how those specific patterns came about remains obscure due to the complexity and lack of transparency of Google’s advertising system.
Datta’s team states that “we cannot determine whether Google, the advertiser, or complex interactions among them and others caused the discrimination. Even if we could, the discrimination might have resulted unintentionally from algorithms optimizing click-through rates or other metrics free of bigotry.” This is an important observation to make: the study does not blame Google directly for creating some inherently biased algorithm, but rather explains that negatively discriminatory algorithmic biases can arise regardless of the intentions of the creator. This is the same issue that occurred with the Amazon hiring algorithm; the developers were unaware of the fact that their data, which seemed harmless, would result in an algorithm biased against women. Thus, tackling the issue of AI/ML bias is not as straightforward as changing the algorithm in question, but rather requires an analysis of the entire system which led to the algorithm’s negative performance. This system can include the data which was used to train the algorithm, the algorithm’s developers, and all third-party entities which are allowed to interact with and influence the algorithm’s results. In the aforementioned Amazon case, the system also includes bias within the English language, showing just how extensively developers must scrutinize their data before using it to train algorithms.
This analysis, however, is not easy to complete without transparency from the organizations in question. Although according to Datta’s team’s work Google Ad Settings does provide transparency in regards to what types of data is being collected, it is unclear how exactly this data is being used. However, some data usage lacked transparency entirely (i.e. was completely opaque) — it was found “that visiting web pages associated with substance abuse changed the ads shown but not the settings page,” meaning there are parts of a user’s advertisement profile which influence their advertisement experience but are not displayed in Ad Settings (Datta et al. 2014). This lack of transparency means researchers could not figure out ways to reduce harmful biases in Google’s advertising algorithm even if they wanted to. Transparency allows users the ability to see how their personal data is being used and protest that usage if they disagree with it. It also allows researchers to access the data which powers the algorithms they study, meaning they can make more informed and beneficial suggestions to help companies like Google improve their algorithms. Since developers may be entirely unconscious of the biases they feed into their algorithms, whether through personal or data bias, it is important that these systems are transparent so users and researchers can point out biases that would otherwise go unnoticed. It is ultimately the responsibility of the company (Google in this example) to ensure that the AI-driven ecosystem that they are developing is not creating any significant disadvantages for a particular set of classes when released for public consumption.
4 Conclusion
Throughout the course of this paper, we have discussed a variety of causes and consequences of bias in AI/ML algorithms. Specifically, we analyzed the algorithmic bias in AI/ML systems through two different lenses: race and gender. Each of the provided examples illustrated how wildly susceptible algorithms are to racial/gender biases, regardless of the data used to train those algorithms and the intentions of the developers. Therefore, it is key to note that impacts on users, rather than developer intentions, should be the point of focus when determining whether or not an AI/ML algorithm possesses ethical or unethical biases.
Latent bias is extremely detrimental and can be classified as unethical when it promotes collective social harm to sensitive protected classes. Collective social harm can be broken down into instances where a class of individuals face a loss of opportunity, economic loss, or social stigmatization due to the biased algorithmic decision-making, which is the case in all of the presented examples. On the other hand, when users’ demographic data is carefully utilized in AI/ML algorithms to create relevant and personalized content, bias in these systems can be deemed ethical. However, despite the fact that some algorithms do implement bias ethically, we argue that AI/ML systems are still too susceptible to negative biases for the current usage of these algorithms to be deemed ethical.
There are some who argue, however, that negative biases in algorithms are not necessarily unethical. This is the case with Northpointe, for example, the company which created the widely-implemented COMPAS algorithm. ProPublica exposed very clear racial inequities in the algorithm’s performance (incorrectly labeling black defendants as high risk for recidivism at a higher rate than white defendants, and incorrectly labeling white defendants as low risk at a higher rate than black defendants) which are mostly indisputable, but as explained by Corbett-Davies et al. from the Washington Post, there is a mathematical basis for Northpointe to object. Furthermore, “Northpointe argues, when judges see a defendant’s risk score, they need not consider the defendant’s race when interpreting it” (Corbett-Davies et al. 2016). This argument is based on the simple fact that the questionnaire used to generate data to be fed into COMPAS does not explicitly ask about a defendant’s race. It also implies that while COMPAS is a tool intended to aid judges in the decision process for criminal sentencing, judges are ultimately responsible for the decision they make. While this is true, if an AI/ML system is flawed or negatively impacts a particular group to a disproportionate extent, it is the developer’s duty to improve that system. It is also somewhat counterintuitive to ask a judge to “not consider the defendant’s race,” as this is basically asking a human to interpret the results of a biased program through an unbiased lens, which is flawed on two fronts: first, humans are fundamentally incapable of being unbiased, and second, if the algorithm was corrected for racial bias, there would be less of a need for the judge to interpret the results in an unbiased manner in the first place.
Nonetheless, the development and implementation of AI/ML systems will no doubt continue to grow pervasively. In order to promote the ethical growth of this technology, it is critical to determine ways to detect and mitigate bias in existing AI/ML algorithms and learn how to create systems that will avoid from the outset undesirable and unintended bias, particularly with respect to SPI. Throughout this paper, a few key issues were identified as factors that allowed for development of detrimentally biased AI: biased developers, prejudiced data, and lack of transparency. These factors are all related to the big-picture issue that while developer and corporate intentions may be benevolent (a lack of transparency does not necessarily mean a company is hiding information), the results of the algorithms they produced may still be negatively and unethically biased due to the widespread discrimination that has existed (and will continue to exist) in society for centuries. However, there are measures that can be taken to address each of these problems. Due to the lack of neutrality that is inherent in humans, it is impossible to have a developer who is completely impartial. By increasing the diversity of individuals working to develop these systems, however, it becomes easier to implement methods to ensure that a training dataset used on a AI/ML system is well balanced. Additionally, while it is important that companies responsibly regulate their own algorithmic development, it cannot be expected that all companies will, and so it is necessary for government bodies to implement comprehensive guidelines detailing standards for how bias in AI/ML algorithms is measured and what kinds of bias qualify as acceptable. This will lead to less controversy, as in the ProPublica vs. Northpointe case, as well as more efficient evaluation of algorithms to ensure that evaluation resources are not wasted and can be applied to improve AI/ML systems elsewhere. Data and algorithmic transparency should also be regulated by government bodies in order to ensure companies are held accountable for unethical systems when they are discovered, and so researchers can access and study data to develop methods for mitigating negative biases. Ultimately, while bias in AI/ML systems can never be removed entirely, and sometimes can be considered ethical and beneficial to society, efforts must be made to keep algorithms with the potential for negative bias operating ethically.
Co-Written with Aaron Rovinsky (Berkeley M.E.T ‘23)
Sources:
Nick Diakopoulos, www.nickdiakopoulos.com/.
“AI Infrastructure Solutions.” IBM, www.ibm.com/it-infrastructure/solutions/ai
Chui, Michael, et al. “What AI Can and Can’t Do (Yet) for Your Business.” McKinsey & Company, www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/what-ai-can-and-cant-do-yet-for-your-business.
Angwin, Julia, et al. “Machine Bias.” ProPublica, 9 Mar. 2019, v2-www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing.
“Artificial Intelligence and Gender Bias.” Catalyst, www.catalyst.org/research/trend-brief-gender-bias-in-ai/.
Bradford, Krista. “Google Shows Men Ads for Better Jobs: Technology Bias.” The Good Search, 1 Dec. 2019, tgsus.com/diversity/google-shows-men-ads-for-better-jobs/.
Dastin, Jeffrey. “Amazon Scraps Secret AI Recruiting Tool That Showed Bias against Women.” Reuters, Thomson Reuters, 10 Oct. 2018, www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G.
Datta, et al. “Automated Experiments on Ad Privacy Settings: A Tale of Opacity, Choice, and Discrimination.” ArXiv.org, 27 Aug. 2014, arxiv.org/abs/1408.6491v1.
Davenport, Thomas H., and Rajeev Ronanki. “3 Things AI Can Already Do for Your Company.” Harvard Business Review, 27 Nov. 2019, hbr.org/2018/01/artificial-intelligence-for-the-real-world.
Larson, Jeff, et al. “How We Analyzed the COMPAS Recidivism Algorithm.” ProPublica, 9 Mar. 2019, www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm.
Ledford, Heidi. “Millions of Black People Affected by Racial Bias in Health-Care Algorithms.” Nature News, Nature Publishing Group, 24 Oct. 2019, www.nature.com/articles/d41586-019-03228-6#ref-CR1.
Obermeyer, Ziad, et al. “Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations.” Science, American Association for the Advancement of Science, 25 Oct. 2019, science.sciencemag.org/content/366/6464/447.
“Project Overview ‘ Gender Shades.” MIT Media Lab, www.media.mit.edu/projects/gender-shades/overview/.
Sam Corbett-Davies, Emma Pierson. “A Computer Program Used for Bail and Sentencing Decisions Was Labeled Biased against Blacks. It’s Actually Not That Clear.” The Washington Post, WP Company, 17 Oct. 2016, www.washingtonpost.com/news/monkey-cage/wp/2016/10/17/can-an-algorithm-be-racist-our-analysis-is-more-cautious-than-propublicas/.
Simonite, Tom. “When It Comes to Gorillas, Google Photos Remains Blind.” Wired, Conde Nast, 20 Nov. 2018, www.wired.com/story/when-it-comes-to-gorillas-google-photos-remains-blind/.
Wills, Craig E., et al. “Understanding What They Do with What They Know.” Understanding What They Do with What They Know | Proceedings of the 2012 ACM Workshop on Privacy in the Electronic Society, 1 Oct. 2012, dl.acm.org/doi/10.1145/2381966.2381969.