Ferrara2023FairnessBias
Literature note on Ferrara, "Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies"
Ferrara, E. (2023). Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies. Sci, 6(1), 3.
Commentary
Ferrara provides an overview of fairness and bias in artificial intelligence applications. In this overview different sources and types of bias are discussed, together with possible real-life implications. Examples of these real-life implications are for example the COMPAS dataset (a dataset used to predict reoffending of criminals) and an AI system that is used in the healthcare domain to predict mortality rates. Examples of the different types of AI bias are representation bias (in which the population is not well represented in the dataset) and measurement bias (in which certain groups are systematically represented too much or too little in the data). Having considered these important aspects of fairness in AI, the author continues to discuss the impact that bias can have in real-world contexts. This section contains the impact bias can have on society as a whole, certain groups and on individuals. Ferrara also discusses the ethical implications of bias and in this section he notes that biased AI systems can negatively impact human autonomy, as it could potentially limit individual freedom. While this is a valid point made by the author, it could have benefitted from more examples or a lengthier explanation. Does this for example only matter for AI systems used within a certain context? However, it is understandable that this is not fully covered in this work, as the main purpose of the paper is to provide an overview of important aspects of fairness and bias in AI.
An interesting addition that this paper provides is the clear overview of ways to mitigate bias in AI systems. Here, Ferrare considers the different stages in which fairness measures can be implemented, and he represents them in a table. This clear presentation of the methods and their advantages and disadvantages provides readers with a way to quickly get a grasp on what kind of method they could use for their own work.
The paper continues with a section on Fairness in AI and a discussion on the difference between fairness and bias. As the two terms are both used in the former part of the paper, I argue that this comparison should be moved to the beginning of the paper. This way, the reader is not confused in the usage of both terms. Besides that, the author once again does a good job providing an overview of different definitions of fairness. The author then continues by describing some mitigating techniques specifically for fairness.
Excerpts & Key Quotes
User bias
Comment:
When discussing the different types of biases, Ferrara discussed user bias as well. The quote above includes his description of this type of bias, and the way in which it can occur. Using this definition of user bias, it seems as if this type of bias could occur with any type of application or method use. A user could also interact with systems in ways that “reflect their own biases”, regardless of whether these systems are AI based or not. Although it is understandable that Ferrara included this bias type in his overview, it still seems confusing for a reader to include the explanation of this type of bias as a more general one.
New types of discrimination
Comment:
In the quote above Ferrara notes the risk of generative AI in creating new types of discrimination or unfairness. This is an interesting notion, because these new types are important to consider in an overview of bias types. However, Ferrara does not include much explanations or examples of these new types of discrimination. An example that I was personally reminded of, can be found in a paper by Peters (2022). In this paper Peters discusses the possible appearance of bias or discrimination based on one’s political preference. While this is usually difficult to detect, an AI system would be able to detect this using proxies or features that humans might not consider. This would likely be an example of new types of discrimination that Ferrara is talking about, but without any further explanation, the reader is left guessing the real meaning.
Peters, U. (2022). Algorithmic political bias in artificial intelligence systems. Philosophy & Technology, 35(2), 25.
- Page 7: “One of the main challenges is the lack of diverse and representative training data"
- Page 7: “Another challenge is the difficulty of identifying and measuring different types of bias in AI systems.”
- Page 7: “Moreover, mitigation approaches may introduce trade-offs between fairness and accuracy.”
- Page 7: “Finally, there may be ethical considerations around how to prioritize different types of bias and which groups to prioritize in the mitigation of bias”
Comment:
In the section containing the quotes above, Ferrara is doing a particularly good job in providing the reader with an overview of the challenges of different bias mitigation techniques. All these considerations are supported by examples or argumentation, which allows the user to make a well thought out decision on what mitigation technique to use. Additionally, these approaches and their challenges are also presented in a clear table overview, which allows the user to quickly get a grasp on the possibilities.