(Unfair) Norms in Fairness Research: A Meta-Analysis

US-centric bias dominates fairness research: The meta-analysis of 139 papers from leading AI ethics conferences (AIES and FAccT, 2018-2022) found that fairness research is overwhelmingly shaped by American perspectives and values, meaning solutions developed may not address fairness concerns relevant to other countries and cultures.

Binary identity categories oversimplify human complexity: The study found widespread reliance on binary classifications like "Black/White" or "male/female" across fairness studies, which fails to capture the nuanced, intersectional nature of human identity and how people actually experience discrimination in real life.

Current fairness definitions may miss real-world harms: By overlooking diverse global contexts and reducing identity to binary categories, fairness research risks creating AI systems that appear fair on paper but fail to address actual bias experienced by people with complex, multifaceted identities across different regions.

Fairness is not objective—it reflects researcher values: The paper emphasizes that fairness in AI is not a technical problem with a single correct answer, but rather a value-laden concept shaped by who is doing the research and what perspectives they bring, making diverse representation in AI research essential.

Systemic change needed in fairness research practices: The authors call for a "paradigm shift" toward more inclusive research design, suggesting the field must actively incorporate global perspectives and nuanced understandings of identity rather than continuing with current approaches.