Redistribution and Rekognition

Abstract

Computer scientists, and artificial intelligence researchers in particular, have a predisposition for adopting precise, fixed definitions to serve as classifiers (Agre, 1997; Broussard, 2018). But classification is an enactment of power; it orders human interaction in ways that produce advantage or suffering (Bowker & Star, 1999). In so doing, it obscures the messiness of human life, masking the work of the people involved in training machine learning systems, and hiding the uneven distribution of its impacts on communities (Taylor, 2018; Gray, 2019; Roberts, 2019). Feminist scholars, and particularly feminist scholars of color, have made powerful critiques of the ways in which artificial intelligence systems formalize, classify, and amplify historical forms of discrimination and act to reify and amplify existing forms of social inequality (Eubanks, 2017; Benjamin, 2019; Noble, 2018). In response, the machine learning community has begun to address claims of algorithmic bias under the rubric of fairness, accountability, and transparency. But in doing so, it has largely dealt with these issues in familiar terms, using statistical methods aimed at achieving parity and deploying fairness ‘toolkits’. Yet actually existing inequality is reflected and amplified in algorithmic systems in ways that exceed the capacity of statistical methods alone. This article outlines a feminist critique of extant methods of dealing with algorithmic discrimination. I outline the ways in which gender discrimination and erasure are built into the field of AI at a foundational level; the product of a community that largely represents a small, privileged, and male segment of the global population (Author, 2019). In so doing, I illustrate how a situated mode of inquiry enables us to more closely examine a feedback loop between discriminatory workplaces and discriminatory systems.

Power and politics in framing bias in Artificial Intelligence policy

“…Intersectional and feminist approaches to AI (Wellner & Rothman, 2020; West, 2020) invite to rethink and challenge discriminatory AI suggesting strategies that are polyvocal, multimodal and experimental (Ciston, 2019) and that promise a fairer, slower, consensual and collaborative AI (Toupin, 2023).…”

Fairness and accountability of AI in disaster risk management: Opportunities and challenges

“…21 Thus, bias can also be the result of algorithm design or decisions ll OPEN ACCESS around metrics used to evaluate a particular phenomenon. While some technical approaches to addressing bias exist, completely eliminating bias in algorithms is impossible, and, as some have argued, 22,23 exclusive focus on reducing bias in AI systems may distract from other, more important, interventions.…”

Practical, epistemic and normative implications of algorithmic bias in healthcare artificial intelligence: a qualitative study of multidisciplinary expert perspectives

“…Of particular, challenge is the inability of AI systems to accurately record and incorporate social data, which include cultural beliefs, economic status, linguistic identity and other social determinants of health 19. While efforts to address algorithmic bias are increasingly interdisciplinary, some scholars argue that the dominant approach of relying on data science mechanisms alone fails to address structural and systemic causes of marginalisation 20–22. Feminist ethicopolitical critiques highlight factors including historically entrenched power structures,22 exclusionary social demographics in the AI workforce,20 21 and sociocultural legacies of colonialism23 as some of the key drivers of algorithmic bias that deserve attention.…”