Imagine the following scenario. Aisha, a talented software engineer with years of experience and a proven track record leading high-performing teams, applies for dozens of leadership positions at major tech companies. She holds a degree from a top university, has led multiple successful product launches, and has fantastic references from colleagues who praise her strategic vision and collaborative style.

Weeks go by, but she doesn’t receive a single interview request. 

Unbeknownst to Aisha, the companies have been using AI-powered hiring tools to filter applicants. The algorithm, trained on historical hiring data, has learned to favor profiles that resemble the company’s past leadership—predominantly male, from a narrow set of colleges and networks. Aisha’s resume includes references to mentoring programs for women in tech and a leadership award from a diversity initiative. Ironically, these markers of excellence trigger a lower ranking in the algorithm’s eyes. As a result, Aisha’s application is quietly filtered out before a human ever sees it.

While the algorithm wasn’t intentionally trying to discriminate, it was replicating the biases embedded in the data it was fed. Over the years, male candidates from certain backgrounds had been favored by hiring teams, leading to a distinct pattern in desired profiles. By replicating this pattern, the algorithm became a gatekeeper that perpetuated hiring inequality. 

Algorithmic fairness is a principle used to assess whether machine learning algorithms operate fairly and if their outcomes are unbiased. That is, to ensure that they don’t unintentionally discriminate against individuals or groups based on characteristics such as race, gender, socioeconomic background, or age.¹ In today’s modern society, algorithmic systems increasingly shape the information we’re exposed to and influence our prospects in areas such as employment, education, and finance.² Algorithmic (un)fairness has also been found to impact child welfare, predictive policing, and online housing marketplaces, prompting urgent conversations about how to prevent injustice in algorithmic systems that influence critical decision-making processes. As such, algorithmic fairness is currently a hot topic for researchers and policymakers alike.⁹

Algorithm: A step-by-step set of instructions or rules for solving a problem or performing a task. Algorithms are a core element of AI processes. 

Machine Learning: A branch of artificial intelligence that enables computers to learn patterns from data and make predictions or decisions without being explicitly programmed.

Demographic parity: In machine learning fairness, demographic parity is a statistical criterion that requires different groups—such as those defined by race, gender, or other protected attributes—to receive positive outcomes at equal rates, regardless of differences in qualifications or behavior. It’s used to check whether a model’s decisions (e.g., loan approvals, hiring) are evenly distributed between groups.

Equalized odds/equal error rates: A fairness definition rule for predictive models that requires each group to have the same false positive rate and the same false negative rate for a given prediction task. In other words, errors are distributed equally across groups, so no one group disproportionately suffers from incorrect decisions. For example, in a medical screening model, the rate of missed diagnoses (false negatives) should be the same for men and women.

Calibration: A performance evaluation method for machine learning models that checks whether the predicted probability matches the actual frequency of an event. For instance, if the model predicts a risk score of 0.7, it would be considered “calibrated” if about 70% of the individuals with the predicted score actually experience the event.

A model is considered ‘calibrated’ if, among all individuals with a predicted risk score of 0.7, about 70% actually experience the event. This relationship must also hold consistently across demographic groups.

Individual fairness: A principle in algorithmic decision-making stating that similar individuals should be treated similarly. “Similarity” is typically defined in terms of relevant attributes for the task. For instance, in a credit scoring model, two applicants with nearly identical financial histories should receive similar scores, regardless of unrelated demographic differences.

Data positivism: The belief or approach that views data as objective, neutral, and inherently truthful.

Algorithmic Accountability Act (AAA): A proposed U.S. federal law that would require companies deploying automated decision systems (ADS) to conduct comprehensive impact assessments for risks such as bias, inaccuracy, or privacy harms. These requirements would apply particularly in domains like health, housing, education, and finance. 

Impossibility theorem: In algorithmic fairness, the impossibility theorem says that if groups differ in how often predictions are right or wrong (their error rates), no model can satisfy all fairness goals at once.

The concept of fairness dates back millennia. As far back as the ancient Greeks, the idea of fairness has been debated, with philosophers and thinkers dedicating hours to finding a true definition. However, the standards we use to judge whether an action or behavior is fair or unfair are highly variable, subjective, and subject to change. Think about death by hanging: while it’s currently outlawed in most countries, for centuries it was seen as a normal and justified punishment for anything from petty theft to murder. Over time, our collective understanding of what’s just and unjust shifts. 

In philosophy and ethics, fairness is determined through a process of deliberation. This idea was first developed by American political philosopher John Rawls, who argued that deciding what counts as “fair” requires reasoned discussion, weighing competing principles, and justifying choices in a way all affected parties could, in principle, agree to.²⁷ Or in less philosophical terms, we say that fairness ‘depends on the circumstances.’⁴ For Greek philosopher Aristotle, fairness was a fundamental condition for human co-existence, necessary for the development of a society that is both enduring and fully realized. In his study of different governments, he said that while there is broad agreement that fairness—or justice—means treating equals the same and unequals differently, people disagree on the criteria for determining who is equal or unequal in terms of worth or deserve.⁵

It may come as a surprise, but algorithms and fairness crossed paths as early as the 17th century, or even before. During the Enlightenment, ideas about ethics played a key role in the early development of probability theory and the concept of mechanical calculation. French mathematician and one of the founders of probability theory, Blaise Pascal, was driven by questions of fairness. He spent a long time deciding how to fairly divide gambling winnings or share profits from risky ventures such as insurance or shipping.⁶ 

By the 19th and early 20th centuries, insurance systems had become more formalized, and the concept of individualized risk (the likelihood that something bad could happen to someone based on their characteristics) emerged. This approach used actuarial tables and statistical models to group people with similar characteristics—such as age, occupation, and health—and assign premiums accordingly. Actuarial tables are statistical charts that show the probability of certain events—like death, illness, disability, or accident—occurring within a given population over a specific period of time. Some insurers also incorporated race as a factor in these calculations. One of the earliest fairness debates arose from racially discriminatory pricing. For example, Black Americans were often charged higher premiums or outright denied life insurance coverage based on actuarial tables that associated race with shorter life expectancy. Insurers claimed this was simply a matter of “risk,” but critics said it built social inequality into economic systems, all while giving it the appearance of being neutral and data-driven.⁶

This practice highlighted a key issue we still see today in algorithmic fairness: the idea that “objective” risk scores or statistical patterns can reflect and perpetuate deep structural bias. The controversy prompted early legal and political efforts to regulate risk classification, foreshadowing modern debates around algorithmic bias in credit scoring, health insurance, and criminal justice.¹⁴

One of the first studies to look specifically at algorithmic fairness was conducted by T. Anne Cleary in 1966.¹² She examined whether standardized tests like the SAT were equally predictive of academic performance across racial groups. Using data from Black and White students at three integrated colleges, Cleary defined a test as biased if it consistently produced prediction errors for members of a particular subgroup. At two of the schools, her analysis found no significant bias. At the third, however, located in the Southwest, the test actually overpredicted the performance of Black students. In that case, the bias favored rather than disadvantaged the subgroup in question, highlighting early on that fairness in predictive systems is not just about accuracy, but about how systematic errors are distributed across different populations.

In the 2010s, research into how algorithmic systems could produce unfair outcomes for marginalized and vulnerable populations accelerated. As more decisions—hiring, credit scoring, bail determinations, school admissions—were either made or supported by automated systems, concerns grew that this outsourcing of decision-making could reinforce or worsen existing social inequalities. Researchers began developing more precise mathematical definitions to address bias in machine learning systems, including demographic parity, equalized odds, calibration, and individual fairness.²⁴ 

According to historian Matthew Jones, the challenge in today’s hyper-digitalized world is “data positivism.”¹³ This is the erroneous belief that data (such as crime statistics, social media analytics, or biometric readings) is objective, neutral, and inherently truthful. That is, data simply “speaks for itself” and reflects reality as it is, without needing interpretation or context. For example, a predictive policing algorithm might treat recorded arrest rates as an unbiased mirror of crime patterns, when in fact those numbers are shaped by historical policing practices and systemic biases. As researchers, civil rights activists, and feminists have warned, bias, discrimination, and injustice exist within our data and algorithmic systems, and we need to be alert to their dangers. 

And it seems that governments and institutions are finally listening. In 2024, the European Union formally adopted the world’s first AI Act (to be enacted in 2026).¹⁵ The purpose of this landmark legislation is to regulate artificial intelligence (before it regulates us) and to ensure that AI systems used within the EU are safe, transparent, fair, and respect fundamental rights. The United States has also proposed similar legislation, first in 2019 and then again in 2023. The Algorithmic Accountability Act (AAA) is a proposed U.S. federal law that aims to bring transparency, oversight, and fairness to automated decision-making systems used by companies—especially those involving artificial intelligence (AI) and machine learning (ML).¹⁶

Blaise Pascal

17th-century French mathematician, physicist, inventor, and philosopher known for his work on probability theory, fluid mechanics, and the invention of one of the first mechanical calculators. In his later years, he turned to religion and wrote the Pensées, which includes his famous philosophical argument, Pascal’s Wager.

Theresa Anne Cleary

Educational psychologist known for her pioneering work in psychometrics and test bias in the 1960s. She held leadership roles at institutions like ETS and the University of Iowa, where she advanced fair assessment practices. Her research significantly shaped how standardized tests are evaluated for equity.

John Rawls

American political philosopher whose work reshaped modern liberal thought. He introduced the concept of justice as fairness, arguing that principles of justice should be chosen under impartial conditions, such as his famous veil of ignorance thought experiment. Rawls’s ideas have had lasting influence on ethics, political theory, law, and debates on fairness in public policy and algorithmic decision-making.

As we allow computers and algorithms to make increasingly important decisions affecting the lives of other people, attention has turned to the unintended impacts that this transfer of power can have. 

Discrimination

Some argue that using algorithms can lead to more objective and potentially fairer decisions than those made by humans, who are often influenced by personal biases and other subjective judgments.¹¹ Yet research has shown that algorithms aren’t perfectly objective tools, but rather reflect the values, assumptions, and historical patterns embedded in the data and design choices behind them. This can lead to the perpetuation of discriminatory decision-making across various sectors, such as hiring, policing, and allocation of social support. Algorithmic fairness reminds us that while they may automate decisions at scale, without careful oversight and transparency, they risk reinforcing existing inequalities under the appearance of neutrality. 

Medical misdiagnosis

Machine learning is increasingly used to help doctors make clinical decisions. AI can analyze large amounts of data quickly, identify patterns, and support faster, more accurate, and more personalized outcomes than humans alone. However, studies have found that these systems can sometimes be biased.¹⁷ As a result, these systems may perform less accurately for individuals in certain demographic groups, potentially exacerbating health disparities and disproportionately impacting the well-being of vulnerable populations. 

For example, common heart disease risk factors don’t always predict outcomes the same way for different ethnic groups. The original models were often developed using data from mostly White populations, so they may not fully capture how risk factors behave in other groups.¹⁸ Similarly, video-based tools designed to track eye blinking may be less effective on Asian individuals.¹⁹ And pulse oximeters, which measure oxygen levels through the skin, have been found to miss low oxygen more often in Black people because darker skin affects how light is absorbed.²⁰ In these cases, technology can create or worsen health inequalities and provide doctors with inaccurate or misleading indicators. 

Erosion of trust

As research on algorithmic fairness highlights the negative impacts of AI bias, some believe that our trust in artificial intelligence systems is eroding.²¹ That is, when algorithmic decisions consistently disadvantage certain groups, public trust in these systems tends to weaken. A group of researchers from various universities in Australia and China believe that algorithmic fairness must be re-examined—not only as a mathematical or technical challenge but as a matter deeply tied to social tolerance and human values.²¹ 

The researchers developed a “tolerance-based fairness” framework that highlights how different stakeholder groups may hold varying thresholds for what they consider fair or acceptable in algorithmic outcomes. Instead of treating fairness as a rigid, one-size-fits-all concept, the authors emphasize the importance of recognizing nuanced social expectations and the contextual nature of fairness perceptions. For example, in a loan approval scenario, the authors explain that risk-averse stakeholders—such as bank executives or regulators—might accept a higher rate of false rejections if it reduces the chance of approving high-risk loans. In contrast, equity-focused stakeholders—such as community advocates—might prioritize expanding access for marginalized applicants and believe that even a small number of unjust denials is a serious violation of fairness.

A tolerance-based approach to fairness can help address discrimination concerns by aligning outcomes more closely with the expectations and values of different communities. Trust, in this sense, goes beyond accuracy or transparency. Instead, it depends on whether people feel the system understands and respects their social identity. 

Fairness in AI is not just a technical challenge but a contested and often political question. Even when scholars and engineers agree on the need to reduce bias, they frequently disagree on what fairness should mean, how it should be measured, and which trade-offs are acceptable. These disagreements fuel some of the most persistent controversies in the field, from theoretical limits on fairness to debates over quick-fix solutions and AI ‘checklists.’

What is fairness?

Since the rise of AI, discussions about fairness have become increasingly prominent. However, the meaning of fairness itself remains deeply contested. One way to define fairness is the absence of bias based on characteristics—whether inherent or acquired—that are irrelevant to the decision at hand. For instance, deciding how much someone should pay for their car insurance based on their gender. But this idea is highly subjective, because what counts as “irrelevant” depends on values, context, and perspective.¹⁰

What makes AI different is not the complexity of fairness, but the ability to automate decisions at scale.³ By relying on past records, historical human judgments, or observed outcomes, AI systems can generate consistent, deterministic decisions without human involvement. Once a model is designed and its input variables fixed, there’s little flexibility to adapt decisions to individual cases. And without proper regulation and oversight, the impacts of this automated decision-making can be huge. 

Today’s supporters of “algorithmic fairness,” often backed by tech companies, usually describe discrimination as a problem that can be solved by balancing different mathematical goals—like accuracy versus fairness.⁶ But they often leave out the bigger challenge: the ongoing political debates about putting legal limits on tools like risk scoring, facial recognition, and drone targeting.

The impossibility theorem

One of the biggest challenges in algorithmic fairness is the so-called impossibility theorem.²⁶ It says that when different groups have different chances of an outcome—like reoffending—no algorithm can be fair to everyone in every way at once. 

Imagine two people with the same risk score: one is a White woman from a wealthier neighborhood, the other a young Black man from a poorer one. For the score to mean the same thing for both, the algorithm has to be “calibrated.” But if one group is more likely to reoffend overall, the error rates—the chance of being wrongly labeled high-risk or low-risk—won’t be equal. You can fix one problem, but then you break the other.

This creates a major dilemma for people building and regulating algorithms. If you focus on reducing one kind of unfairness, like making sure groups get equally accurate results, you might end up sacrificing another, like making sure that risk scores are equally meaningful across groups. The impossibility result shows that fairness in AI is about making value-based decisions and trade-offs, depending on what kind of harm you’re trying to prevent. This insight has sparked a lot of debate, and it’s one reason why fairness in algorithms remains such a hard and deeply political issue.

The limits of checklists

Another debate in the world of algorithmic fairness is whether we can use simple tools like checklists or automated tests to find and fix unfair algorithms. Many engineers like the idea of having a clear step-by-step process—something they can follow to quickly spot problems. But others, especially social scientists, argue that fairness is too complicated for a one-size-fits-all solution.²⁵

What’s fair often depends on the specific situation, the people involved, and the context. A checklist might catch some issues, but it could also miss deeper problems or give a false sense of confidence. While some useful guidelines might emerge over time, experts warn against relying too much on quick fixes. The good news is that more people across different fields are starting to recognize that fairness isn’t just a technical issue; it’s also social and political. One such example is FAT*, now known as FAccT (Fairness, Accountability, and Transparency): a multi-disciplinary group that takes a holistic approach to the problem, pushing research beyond just technological solutions. 

COMPAS

COMPAS, or Correctional Offender Management Profiling for Alternative Sanctions, has been at the center of heated debate within the US criminal justice system for over a decade. The tool is designed to rate defendants based on how likely they are to commit another crime if released. This information could be useful for judges and parole boards, who are often required to make complex decisions about bail, sentencing, or release. 

COMPAS looks at factors such as a person’s job, housing, personality traits, and criminal record to calculate how likely they are to commit another crime. It gives a score from 1 (very low risk) to 10 (very high risk), based on how that person compares to a large group of past offenders with similar profiles. In other words, the model looks at what people with similar backgrounds and histories have done in the past, and uses that to predict future risk. Judges often take advantage of that score in deciding things like bail, parole, or prison sentence. If someone gets a high score by mistake (a false positive), it can lead to a tougher sentence or being denied release. But if they get a low score by mistake (a false negative), it usually helps them.

However, in 2016, investigative journalism group ProPublica analyzed the data produced by COMPAS, comparing the algorithm’s predictions of reoffending with actual reoffense records over a two-year period. They found that while it was accurate for making correct predictions (i.e., flagging someone who actually goes on to reoffend), it was biased when making false predictions. COMPAS, the authors argued, is much more likely to falsely flag Black defendants who won’t reoffend as high risk (45%) than White defendants (24%).⁸ At the same time, it is more likely to wrongly label White people who will reoffend as low-risk (about 48%) than Black people (about 28%).³ This is highly problematic because false decisions around recidivism—the tendency to relapse into criminal behavior—can significantly impact a person’s life and their future prospects. 

So how did the owners of COMPAS respond to these allegations? Northpointe, the company that sells the tool to US courts, said that the machine was “well-calibrated.” This means that if two people, one Black and one White, both receive the same risk score, they are equally likely to reoffend. In other words, the score works the same way across groups. Some experts, including the folks at Northpointe, argue that this kind of consistency makes the system fair, even if other measures—like false positive or false negative rates—differ between groups.⁸

However, calibration and equalized error rates cannot both be achieved when groups have different base rates of reoffending. In other words, COMPAS can be calibrated across groups, but this will necessarily come at the expense of equalized odds, and vice versa. This mathematical incompatibility is at the heart of the fairness debate.

LinkedIn’s AI-bias algorithm

LinkedIn has fast become the world’s go-to social media platform for anything career or job-related. Like many online platforms, LinkedIn relies on algorithms to help users navigate the enormous amount of content available. For instance, the “People You May Know” (PYMK) algorithm suggests potential connections to help users expand their professional networks. The InstaJob algorithm alerts job seekers to new postings they might be interested in, while the Recruiter Search algorithm ranks and displays the most relevant candidates when a recruiter performs a keyword search.

However, the platform’s AI-powered job recommendation systems have been criticized for reinforcing gender bias. In 2016, the Seattle Times reported that LinkedIn’s search algorithm sometimes suggested male names when searching for female ones.²³ For example, offering “Stephen Williams” as a result when searching for “Stephanie Williams.” However, when looking up common male names in the U.S., the search algorithm didn’t suggest female names. LinkedIn tried to brush the problem off by suggesting that the algorithm generated results based on users’ past search tendencies. But clearly, there was a gender bias problem. 

A few years later, in 2022, a study by LinkedIn employees introduced a fairness metric to explore this issue and found that the platform’s algorithms were more likely to favor male candidates over equally qualified female candidates, resulting in uneven job recommendations.²² When comparing similarly qualified profiles (matched by experience, education, skills, etc.), men were more likely to appear at the top of ranked lists. And since higher-ranked candidates are more likely to be contacted or hired, this disparity can lead to reduced job opportunities for women. 

However, it’s not all the algorithm’s fault. While the platform relies on various algorithms to shape user experiences through personalized recommendations, these outcomes are also influenced by how users interact with one another on the platform. In the context of gender bias, this means that even if LinkedIn’s algorithms are designed to be fair, they may still learn and reinforce biased patterns that emerge from user behavior—such as who gets clicked on, followed, or recommended. In other words, the algorithm learns human biases inherent in the job market. Ensuring fairness, then, requires acknowledging that not all sources of inequality can be solved through technical fixes alone, especially when they stem from broader social dynamics beyond the platform’s control.

AI algorithms at work: How to use AI to help overcome historical biases

AI, often seen as a source of bias, can actually be used to counteract it. By anonymizing inputs and focusing on behaviorally relevant data, AI can reduce reliance on stereotypes and irrelevant demographic cues in decision-making. In this article, Ariel Lafayette and Turney McKee argue that when designed with behavioral science principles, AI has the potential to expose and correct deeply embedded structural inequities.

Ethical AI

Algorithmic fairness is part of a broader concept of ethical AI. In this article, Annika Steele explores how ethical AI involves designing systems that prioritize fairness, accountability, transparency, and human rights. Algorithmic fairness is just one component in ensuring AI promotes social good rather than reinforcing existing inequalities.

1. Pophal, L. (n.d.). What is algorithmic fairness? Visier HR Glossary. Retrieved August 2, 2025, from https://www.visier.com/blog/hr-glossary-what-is-algorithmic-fairness/
2. Woodruff, A. (2019). 10 things you should know about algorithmic fairness. interactions, 26(4), 47–51. https://doi.org/10.1145/3328489
3. Pfeiffer, J., Gutschow, J., Haas, C., & others. (2023). Algorithmic fairness in AI. Business & Information Systems Engineering, 65(3), 209–222. https://doi.org/10.1007/s12599-023-00787-x
4. Weinkauf, D. (2023, April 5). When worlds collide – The possibilities and limits of algorithmic fairness (Part 2). Privacy Tech‑Know Blog. Office of the Privacy Commissioner of Canada. Retrieved August 2, 2025, from https://www.priv.gc.ca/en/blog/20230405_02/
5. Aristotle. (1998). Politics (C. D. C. Reeve, Trans.; Book III, Chap. 9). Hackett Publishing. (Original work published ca. 350 BCE)
6. Ochigame, R. (2020, January 30). The long history of algorithmic fairness. Phenomenal World. Retrieved August 2, 2025, from https://www.phenomenalworld.org/analysis/long-history-algorithmic-fairness/
7. Bosavage, J. (2019, February 25). 10 quotes about artificial intelligence from the experts. Oracle. https://blogs.oracle.com/cx/post/10-quotes-about-artificial-intelligence-from-the-experts
8. Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016, May 26). Machine bias: There’s software used across the country to predict future criminals. And it’s biased against Blacks. ProPublica. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
9. Huang, Y., Liu, W., Gao, W., Lu, X., Liang, X., Yang, Z., Li, H., Ma, L., & Tang, S. (2023). Algorithmic fairness in social context. BenchCouncil Transactions on Benchmarks, Standards and Evaluations, 3(3), Article 100137. https://doi.org/10.1016/j.tbench.2023.100137
10. Saxena, N. A., Huang, K., DeFilippis, E., Radanovic, G., Parkes, D. C., & Liu, Y. (2019). How do fairness definitions fare? Examining public attitudes towards algorithmic definitions of fairness. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society (pp. 99–106). ACM. https://doi.org/10.1145/3306618.3314248
11. Lepri, B., Oliver, N., Letouzé, E., Pentland, A., & Vinck, P. (2017). Fair, Transparent, and Accountable Algorithmic Decision-making Processes. Philosophy & Technology, 31, 611 - 627. https://doi.org/10.1007/s13347-017-0279-x.
12. Cleary, T. A. (1966). Test bias—Validity of the Scholastic Aptitude Test for Negro and White students in integrated colleges. ETS Research Bulletin Series, 1966(2), i–23. https://doi.org/10.1002/j.2333-8504.1966.tb00192.x
13. Jones, M. L. (2018). How we became instrumentalists (again): Data positivism since World War II. Historical Studies in the Natural Sciences, 48(5), 673–684. https://doi.org/10.1525/hsns.2018.48.5.673
14. Levy, J. (2012). Freaks of fortune: The emerging world of capitalism and risk in America. Harvard University Press.
15. European Commission. (2021). Proposal for a Regulation of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52021PC0206
16. Wyden, R., Booker, C., & Clarke, Y. (2022). Algorithmic Accountability Act of 2022 (U.S. Senate Bill S.3572). https://www.congress.gov/bill/117th-congress/senate-bill/3572
17. Xu, J., Xiao, Y., Wang, W. H., Ning, Y., Shenkman, E. A., Bian, J., & Wang, F. (2022). Algorithmic fairness in computational medicine. eBioMedicine, 84, Article 104250. https://doi.org/10.1016/j.ebiom.2022.104250
18. Gijsberts, C. M., Groenewegen, K. A., Hoefer, I. E., Eijkemans, M. J. C., Asselbergs, F. W., Anderson, T. J., ... & den Ruijter, H. M. (2015). Race/ethnic differences in the associations of the Framingham risk factors with carotid intima-media thickness and cardiovascular events. PLoS ONE, 10(7), e0132321. https://doi.org/10.1371/journal.pone.0132321
19. Zou, J., & Schiebinger, L. (2018). AI can be sexist and racist—it’s time to make it fair. Nature, 559(7714), 324–326. https://doi.org/10.1038/d41586-018-05707-8
20. Sjoding, M. W., Dickson, R. P., Iwashyna, T. J., Gay, S. E., & Valley, T. S. (2020). Racial bias in pulse oximetry measurement. The New England Journal of Medicine, 383(25), 2477–2478. https://doi.org/10.1056/NEJMc2029240
21. Luo, R., Tang, T., Xia, F., Liu, J., Xu, C., Zhang, L. Y., Xiang, W., & Zhang, C. (2024). Algorithmic fairness: A tolerance perspective. arXiv preprint arXiv:2405.09543.
22. Yu, Y., & Saint-Jacques, G. (2022). Choosing an algorithmic fairness metric for an online marketplace: Detecting and quantifying algorithmic bias on LinkedIn. arXiv preprint arXiv:2202.07300. https://arxiv.org/abs/2202.07300
23. Day, M. (2018, October 12). How LinkedIn’s search engine may reflect a bias. The Seattle Times. https://www.seattletimes.com/business/microsoft/how-linkedins-search-engine-may-reflect-a-bias/
24. Hardt, M., Price, E., & Srebro, N. (2016). Equality of opportunity in supervised learning. In Advances in Neural Information Processing Systems (NeurIPS), 29, 3315–3323. https://proceedings.neurips.cc/paper_files/paper/2016/file/9d2682367c3935defcb1f9e247a97c0d-Paper.pdf
25. Woodruff, A. (2019). 10 things you should know about algorithmic fairness. interactions, 26(4), 47–51. https://doi.org/10.1145/3328489
26. Sahlgren, O. (2024). What’s impossible about algorithmic fairness? Philosophy & Technology, 37, Article 124. https://doi.org/10.1007/s13347-024-00814-z
27. Rawls, J. (1971). A theory of justice. Harvard University Press.