Data Feminism for AI

Lauren Klein and Catherine D’Ignazio do not turn the seven principles of Data Feminism into a checklist for scoring finished models. This theoretical paper, presented at FAccT in 2024, explains why data feminism remains necessary to understanding AI: the issue is not only whether a model is biased, but who sets research agendas, whose data are collected and interpreted, who owns the infrastructure, and how a system’s benefits and costs are distributed. The authors use intersectionality as a way to analyze structural power, not as a request to add more identity labels to a model. They restate the seven principles from Data Feminism for AI; environmental impact and consent are two possible directions for further development, not finalized additions to the framework. The official conference PDF gives the proceedings citation and conference dates, June 3–6, 2024 (print pp. 100–112).

“Examine power” becomes concrete in the discussion of US tenant-screening services. These systems combine eviction records, arrest records, and credit scores to predict a tenant’s ability to pay rent. The underlying records are inconsistent and are produced within a deeply racialized housing system, so automated scoring can recast existing exclusion as neutral prediction. The authors point to an information imbalance: landlords can use data to know a great deal about tenants, while tenants often lack tools to investigate unfair evictions or neglected properties. A feminist design question therefore shifts from helping landlords screen more efficiently to imagining systems that build tenants’ collective knowledge and power (print pp. 102–103). Better accuracy by itself may leave untouched whom a system serves and which institution it strengthens.

The paper’s account of situated knowledge and embodiment asks whose warnings count as knowledge. Klein and D’Ignazio revisit Joy Buolamwini’s audit of facial recognition using her own face, and research led by Black women documenting misogyny, racism, and sexual violence in the LAION multimodal datasets. Lived experience is not treated as noise that contaminates objectivity; it can provide a crucial standpoint for detecting harms that dominant evaluation misses. The authors also examine how authority is distributed: they note that concerns from Black women-led research received greater institutional attention after white researchers at Stanford replicated the findings (print pp. 104–105). “Representation” here means more than diverse samples. It also concerns who is authorized to interpret evidence and whose judgment organizations recognize.

A further example comes from a news classifier co-designed with feminist data activists across the Americas. The project initially aimed to use machine learning to identify news about feminicide and reduce the activists’ sorting workload. The activists resisted full automation: definitions of feminicide could not simply be generalized across contexts, and documenting those killed was also a practice of memory justice, not merely a task to process faster. Co-design thus did more than solicit feedback before launch; it changed the tool’s purpose and the boundary of automation. The paper reports that the resulting tool was later taken up by human-rights monitoring groups (print pp. 105–106). The sharpest question is not simply whether AI can classify, but who can refuse when political memory is converted into a scalable workflow—and whether that refusal changes the design.

“Make labor visible” moves attention from the apparent intelligence on screen to the people whose work sustains it. Data annotation and content moderation often occupy the bottom of technical job hierarchies, with status and compensation patterns the authors connect to gendered, racialized, and colonial divisions of labor. They cite reporting after ChatGPT’s release about workers in Kenya screening potentially harmful outputs in real time (print pp. 107–108). This is an existing case used to develop the paper’s argument, not a new labor study conducted by the authors. It makes “automation” a political description: what disappears from view is not human work itself, but its conditions, pay, and workers’ power to speak about it.

The discussion treats environmental impact and consent as subjects for further development. The authors place AI’s water and electricity use within longer histories of capitalist and colonial resource extraction, arguing that environmental costs fall unevenly across geography and gender. They also question whether consent understood as an individual, binary, one-time permission can address training-data provenance, intimate images synthesized without consent, or relationships of mutual dependence. They suggest further consideration of collective, interdependent, and queer forms of consent (print pp. 108–109). These are discussions and a research agenda, not operational standards already supplied by the paper. Their importance is to expose infrastructure costs and relational responsibilities that a “fair model” score may leave out.

The paper’s limits matter too. It is a theoretical synthesis of prior research, examples, and principles, not an empirical evaluation proving that one set of principles makes AI systems equitable. Its cases arise from specific political and geographic contexts and should not be generalized into a claim that all communities experience AI in the same way. For a reading path, pair it with Data Feminism to revisit the seven principles’ foundations, then Feminist Data Visualization to see how the same authors bring feminist epistemology into representation. Read the FAccT conference paper; its bibliographic record and DOI are available through the ACM Digital Library.