Ninareh Mehrabi et al., "A Survey on Bias and Fairness in Machine Learning"

Bibliographic info

Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., & Galstyan, A. (2021). A Survey on Bias and Fairness in Machine Learning. ACM Computing Surveys (CSUR), 54(6), 1–35. https://doi.org/10.1145/3457607

Commentary

What I appreciated most about this survey is that it shows how bias in machine learning is not limited to a single stage of the development process. Instead, the authors explain how bias can emerge during data collection, model design, deployment and even through user interaction after a system has been released. This broader perspective helped me see fairness as a continuous process rather than something that can be "fixed" by choosing the right algorithm.

I also found the paper useful because it provides a clear overview of the many definitions of fairness that exist in the literature. Before reading it, I tended to think of fairness as a single objective that could be measured mathematically. The survey makes it clear that different fairness metrics reflect different values and priorities, and that choosing one metric over another is not purely a technical decision. This connects well with discussions from the Digital Ethics course, where we often saw that ethical questions cannot be solved by technical optimisation alone.

One limitation of the paper is that, although it presents an extensive overview of fairness metrics and mitigation techniques, it spends relatively little time discussing the broader social and political context in which these systems operate. The survey explains how to reduce different forms of bias, but gives less attention to questions about power, historical discrimination or who ultimately decides what counts as "fair". As a result, I think the paper works best when complemented by more philosophical or sociotechnical perspectives on AI ethics.

Excerpts & Key Quotes

Bias as a feedback loop

• Page 115:3 (Passage 5)

"Furthermore, we observe that biased algorithmic outcomes might impact user experience, thus generating a feedback loop between data, algorithms, and users that can perpetuate and even amplify existing sources of bias."

Comment

This was probably the most important idea in the paper for me. It shows that bias is not simply something that exists in historical datasets. Once an AI system is deployed, its predictions influence how people behave, and these new behaviours become future training data. In other words, bias can reinforce itself over time. This makes continuous monitoring and accountability just as important as building a good model in the first place.

Fairness has no universal definition

• Page 115:11 (Passage 45)

"The fact that no universal definition of fairness exists shows the difficulty of solving this problem. Different preferences and outlooks in different cultures lend a preference to different ways of looking at fairness, which makes it harder to come up with just a single definition that is acceptable to everyone in a situation."

Comment

I found this quote particularly interesting because it reminds us that fairness is not only a mathematical concept but also a social one. Engineers often look for a single "correct" metric, but this paper shows that different contexts may require different definitions of fairness. Choosing a fairness metric therefore involves value judgments, not only technical decisions. This idea connects closely with our discussions about justice, ethics and the importance of considering different stakeholders when designing AI systems.

Equality versus equity

• Page 115:25 (Passage 93)

"The definitions presented in the literature mostly focus on equality, ensuring that each individual or group is given the same amount of resources, attention, or outcome. However, little attention has been paid to equity, which is the concept that each individual or group is given the resources they need to succeed."

Comment

This distinction between equality and equity was one of the most interesting parts of the survey. Many technical fairness metrics focus on treating everyone equally, but this may not always produce fair outcomes if different groups start from unequal circumstances. Treating everyone the same is not necessarily the same as treating everyone fairly. I think this is an important reminder that improving AI systems requires understanding the social context behind the data, not only improving model performance.

bias
fairness
algorithmic bias
accountability
justice
ethics
machine learning
equity