TLDR: This research paper introduces “Substantive Intersectional Algorithmic Fairness,” a new framework that integrates intersectional feminist theory into the design and deployment of AI systems. It critiques current algorithmic fairness approaches for oversimplifying social realities and proposes ten desiderata, organized under the ROOF methodology, to guide practitioners in addressing systemic inequities, acknowledging power dynamics, and ensuring fairness is deeply rooted in social context, even suggesting non-deployment when necessary.
The rapid advancement of Artificial Intelligence (AI) has brought about significant societal changes, but also raised critical questions about fairness, equity, and discrimination. While the field of algorithmic fairness has emerged to address biases in AI systems, a new research paper argues that current approaches often fall short by oversimplifying complex social realities. Titled “A Feminist Account of Intersectional Algorithmic Fairness,” this paper introduces a groundbreaking framework that integrates intersectional feminist theory to create more equitable and context-sensitive algorithmic practices.
Understanding Intersectionality in AI
At the heart of this research is the concept of intersectionality, a term coined by U.S. legal scholar Kimberlé Crenshaw in 1989. Intersectionality reveals how interconnected systems of privilege and oppression, such as racism, sexism, and classism, jointly shape individuals’ lived experiences. Historically, movements for civil rights and feminism often failed to account for the compounded forms of oppression faced by those who were simultaneously marginalized, like African-American women. The paper highlights how this “single-axis lens” on discrimination obscured the multifaceted ways in which overlapping systems of oppression operate.
In the context of AI, existing algorithmic fairness methods frequently adopt a similar single-axis framework, analyzing fairness through isolated attributes like race or gender. Even efforts to account for intersectional biases often reduce it to auditing and adjusting for subgroup disparities (e.g., gender × race), which the authors argue computationally reproduces the very issues feminist activists have been fighting against for decades. Intersectionality, as the paper emphasizes, is not just about using the term, but adopting a way of thinking about power, sameness, and difference that acknowledges the system-level nature of discrimination and the diversity of lived experiences.
Introducing Substantive Intersectional Algorithmic Fairness
The researchers propose “Substantive Intersectional Algorithmic Fairness,” an extension of the notion of substantive algorithmic fairness, enriched with insights from intersectional feminist theory. This approach fundamentally argues that fairness cannot be separated from its social context. It moves beyond merely identifying biases to actively addressing systemic inequities and mitigating harms to multiply marginalized communities.
To guide this new approach, the paper introduces ten desiderata, or minimal requirements, organized under the “ROOF methodology”: Recognition, Overcoming, Overcoming, and Ways Forward. These desiderata are not fixed operationalizations but encourage deep reflection on the assumptions underlying algorithmic systems.
Key Desiderata of the ROOF Methodology
The ROOF methodology provides actionable guidance for designing, assessing, and deploying AI systems:
- Recognition of Basic Epistemological Assumptions: This involves questioning the assumed neutrality of decision processes, making the positionality of researchers and practitioners explicit, and using precise language that explicitly addresses oppression rather than euphemisms like “algorithmic bias.” It acknowledges that all knowledge is situated and that AI systems embody assumptions reflecting the worldviews of their creators.
- Overcoming the Narrow Focus on Protected Subgroups: This means critically examining the meaning of social categories used in datasets, recognizing them as historically contingent and politically constructed, rather than static labels. It also emphasizes that a substantive approach does not weigh or order different forms of oppression, resisting the temptation to prioritize categories based on data availability or statistical significance.
- Overcoming the Lack of Attention to Socio-Technical Systems: This section calls for mapping power and domination structures within which algorithms operate. It acknowledges that even small algorithmic actions can have significant and distinct impacts on different social groups, similar to microaggressions. Furthermore, it stresses the importance of aligning the purpose of an algorithmic system with its social context and actual impact, and explicitly considering privileges, not just disadvantages, to understand the full scope of social hierarchies.
- Enhancing Ways Forward: Finally, the framework encourages recognizing the opportunity for algorithmic systems to go beyond mere critique and serve as tools for social transformation. This means deploying AI not just to diagnose discrimination but to actively shape equitable decision-making infrastructures and deliver tangible benefits to marginalized groups.
Critiquing Current Approaches
The paper thoroughly critiques existing “Formal Intersectional Algorithmic Fairness” approaches, highlighting several shortcomings. These include an over-reliance on statistically significant subgroups, which often neglects underrepresented communities, and a focus on mitigating disadvantage without examining systemic privilege. Critics also point out that current methods often detach social categories from their historical and power contexts, ignore the real-world consequences of algorithmic decisions, and frame discrimination as individual misconduct rather than a product of deeply embedded social structures. There’s also a noted lack of reflexivity among researchers regarding their own role in reproducing inequalities.
A Call for Principled Non-Deployment
A crucial aspect of Substantive Intersectional Algorithmic Fairness is the recognition that in some cases, principled non-deployment of an algorithmic system may be necessary. This acknowledges that not all social problems require technical solutions, and sometimes, the potential for harm outweighs any perceived benefits, especially for multiply marginalized groups.
Also Read:
- Navigating AI’s Moral Compass: A Call for Dynamic Value Alignment
- Uncovering Hidden Biases: How LLMs Infer Demographics from Neutral Questions
Bridging Disciplines for a More Just Future
By bridging computational and social science perspectives, this research provides actionable guidance for more equitable, inclusive, and context-sensitive intersectional algorithmic practices. It challenges the machine learning field to move beyond techno-solutionism and embrace a deeper understanding of social justice, emphasizing that true fairness requires a critical examination of power, context, and the lived experiences of all people. For more in-depth information, you can read the full research paper here.
