Youjin Kong
This paper critically analyzes the notion of intersectionality employed in research on fairness in artificial intelligence (AI), and thereby aims to contribute to building fairer AI algorithms.
With the proliferation of studies on (un)fairness in AI, recent years have seen emerging interest specifically in “intersectional” bias and unfairness in AI. For example, face recognition algorithms tend to perform better on recognizing men than recognizing women (gender discrimination) and better on people with lighter skin tones than people with darker skin tones (race discrimination); they yield much worse accuracy on women with darker skin tones (intersection of gender and race discrimination).
Most studies in computer science that use intersectionality as a framework for measuring fairness take “intersectional fairness” to be a matter of achieving parity across intersectional subgroups. According to this view, an AI algorithm is fair if the probabilities of outcomes (e.g., being hired, being approved for a loan) are the same or similar regardless of a group’s intersecting combination of attributes such as gender and race—that is, if the probabilities are equal between all subgroups with different combinations of these attributes.
In this paper, I identify and examine four problems with this dominant interpretation of intersectional fairness.
First, the dominant approach is so preoccupied with the intersection of identity categories (e.g., race, gender) that it fails to address the intersection of oppression (e.g., racism, sexism), which is more central to intersectionality as a critical framework. For example, black women are oppressed not because they have intersecting identities of “black” and “women” per se, but because these identities are shaped by and lived in the intersecting structure of racism and sexism. The exclusive focus on identities/attributes may divert attention from structural oppression that causes unfairness between subgroups.
Second, getting the same probabilities of outcomes between subgroups serves “mechanical equality” rather than “fairness.” The dominant approach fails to recognize that, in order to undo the effects of structural oppression and make AI-driven decision-making fairer, more active interventions (such as affirmative action) that allow higher probabilities of positive outcomes for marginalized subgroups than for privileged subgroups are needed.
The third and fourth problems pertain to how AI algorithms work. There is a myriad of factors that lead to a certain outcome from an AI algorithm (e.g., hiring algorithm: applicants’ race, gender, immigration status, first language, age, education level, disability, and so on).
On the one hand, if the algorithm seeks to get equal probabilities across all possible intersecting combinations of these factors, it would keep splitting groups into finer subgroups until the individual is the only cohesive unit of analysis (infinite regress).
On the other, if researchers single out only a couple of factors (e.g., race, gender) to measure fairness, this would be a case of arbitrary selection of a few axes of oppression to pass the fairness test, as opposed to consideration of the entire pattern of intersecting oppressions (“fairness gerrymandering”).
I conclude by suggesting ways to better implement intersectionality as a critical framework for improving fairness in AI.
Feminism, Social Justice, and AI 