Taking a critical look at the critical turn in data science: From “data feminism” to transnational feminist data science

Abstract

Through a critical analysis of recent developments in the theory and practice of data science, including nascent feminist approaches to data collection and analysis, this commentary aims to signal the need for a transnational feminist orientation towards data science. I argue that while much needed in the context of persistent algorithmic oppression, a Western feminist lens limits the scope of problems, and thus—solutions, critical data scholars, and scientists can consider. A resolutely transnational feminist approach on the other hand, can provide data theorists and practitioners with the hermeneutic tools necessary to identify and disrupt instances of injustice in a more inclusive and comprehensive manner. A transnational feminist orientation to data science can pay particular attention to the communities rendered most vulnerable by algorithmic oppression, such as women of color and populations in non-Western countries. I present five ways in which transnational feminism can be leveraged as an intervention into the current data science canon.

Introduction

Keeping a watchful eye on the profound technological innovations of today is a growing body of critical data and algorithm studies (CD&AS) literature, which seeks to critically elucidate the effects of algorithms on society and the environment. The field spans the epistemological gamut from critical practitioner analysis of the limits of technology (Broussard, 2018; O’Neil, 2016), through post-Marxist class-based critique (Eubanks, 2018; Zuboff, 2019), to critical race theory (Atanasoski and Vora, 2019; Benjamin, 2019). The fact that the cited authors are women cannot be underestimated in the face of a computational industry that has historically—and alarmingly—favored white men. The staying power of these women's insights and their inspiring commitment to justice serve as the foundation of data feminism. Among feminist CD&AS works, Catherine D’Ignazio and Lauren Klein's 2020 book Data Feminism stands out with its roadmap for feminist data science.

Mobilizing a transnational feminist analysis, the goal of this commentary is twofold: (1) to read the field of data feminism embodied by D’Ignazio and Klein's project transnationally and point to several conceptual blind spots which limit the field's stated aims, and (2) to underscore the importance—and, indeed, the need— to center transnational feminist interventions in data science given the global impact of algorithms. To accomplish this dual objective, I propose five ways in which a transnational feminist lens can be mobilized in the study and practice of data science, exemplified by CD&AS works that have already begun to successfully leverage such a lens.

Towards a transnational feminist critique of data science

• 1.

  The Importance of a Non-essentializing Feminist Framework

At the crux of data feminism is the development of a framework for practicing data science based on feminist principles such as naming and resisting power in data science projects and producing emotionally evocative graphs (D’Ignazio and Klein, 2020). The way the authors strive to scaffold the authentic nature of feminist data science is reminiscent of Hélène Cixous’ project of conceptualizing the essence of “feminine writing”—écriture feminine—as part of feminist theory (Castle, 2009). Capturing the “true essence” of feminine writing however, has been rightfully critiqued as overly essentialist (Butler, 2002) and a similar critique can be offered against the feminist data science framework proposed by D’Ignazio and Klein.

In light of this critique, insights from postcolonial studies and transnational feminism can be used to re-orient Data Feminism's quest for universals and essentializing authenticity to thinking about ways (necessarily plural) of doing data science differently from the mainstream dominant racist, capitalist, and patriarchal regime. For example, a transnational feminist orientation to data science would ask us to consider the avenues available for forming politically effective communities of data science theorists and practitioners rather than pursuing this work via a collection of disconnected individuals. Such a communal view is indispensable to data science because it centers the constant flux and evolution of political struggles rather than the elusive fixity of a mythical set of essential feminist principles. A focus on the political and transnational underpinnings of data and technology as exemplified by the work of feminist CD&AS scholars such as Avrina Jos, Eden Medina, Radhika Gajjala, Miglena Nikolchina, and Ewa Ziarek (Medina, 2011; Gajjala, 2004; Jos, 2021; Nikolchina, 2021; Ziarek, 2020), rather than a prescription for a singular ‘proper feminist enactment,’ enables women, especially non-Western women, to relate to the field's history and chart out collective strategies for disrupting injustice in an interactive and provisionary manner, and not according to a preset agenda. This is important since, as transnational feminist praxis shows, the structures of oppression are dynamic and constantly changing; consequently, prescribed precepts have limited applicability and effectiveness (Mohanty, 2003).

• 2.

The Interconnectedness of Data Science Struggles

Another contribution of transnational feminism is the tenet that an effective form of resistance is reflecting on how our localized set of conditions relates to global patterns. It reminds us that what may seem like a collection of isolated struggles specific to a given context in fact constitutes an interdependent web such that the struggle of Black Lives Matter activists in the United States, for example, is not disconnected from the fight for freedom in Gaza (Davis, 2016). Such an expanded transnational view makes it possible to understand the connections between historically, geographically, and socially distant struggles and form an effective basis to resist harmful forces that are ultimately threatening to all communities. It also enables us to think beyond the present moment in an active exercise of political imagination about what the world could look like. A commitment to Indigenous futurity (Tuck and Gaztambide-Fernández, 2013), for instance, is a crucial consideration which can help illuminate urgent questions such as the role of data science in the current ecological crises we all face.

This spatially, temporally, and thematically inclusive standpoint can be of tremendous value to the study of data science by helping to connect the dots between seemingly unrelated phenomena such as the almost total surveillance of Uighurs in Xinjiang, Kremlin-sponsored “troll farms,” facial recognition proliferation in the United States and United Kingdom, and political suppression in Colombia, among many others. Preeminent CD&AS scholars like Zuboff (2019) conduct deep analyses of such computationally-enabled repressive systems, but they do so in a localized fashion. Although a class-inflected historicist analysis of the kind Zuboff (2019) presents is a much-needed intervention in data science, it remains at its core US-centric, creating the impression that the algorithmic logics employed by Big Tech in the United States are somehow fundamentally different from the technological workings of nationalized surveillance systems such as China's, when in fact both rely on the common principles of information warfare, albeit for ostensibly different ends (Ruhmann and Bernhardt, 2019). The transnational orientation of important CD&AS research areas such as postcolonial computing (Irani et al., 2010), Big Data from the South(s) (Arora, 2016; Milan and Treré, 2019), and computing from the South (Amrute and Murillo, 2020) overcomes this artificial separation and helps highlight the dichotomized axes of algorithmic power which transcend national borders despite some undeniably nationally-specific characteristics and effects.

• 3.

Beyond the Western Canon of the Enlightenment

Transnational feminist praxis also makes known the existence of plural coexisting worlds rather than typifying the experiences of a discursively preconstituted class of people (Mohanty, 2003) as in the globalist discourse of information and communication technologies. Prioritizing the Western experience in order to create a homogenized, monocultural world serves to totalize Western data science as the only (or certainly the most legitimate) body of knowledge available—and it does so at the expense of the developing world, whose experiences may not fit within this tidy epistemological narrative. A transnational view, however, would make legible epistemologies that are neglected by the hegemonic data scientific canon, such as the Middle Eastern practice of jidal, or debate, described by Fatima Mernissi. Mernissi (2009) points to the possibilities of the Internet, which enables ordinary people previously unengaged in the public debate of democracy, to voice their opinions and address each other in the debate process. This conceptualization of online interactions stands in noticeable contrast to Western-centric theoretical preoccupations with the way Internet users are influenced by online content. Even studies of the radicalization of large groups of people, such as those behind the 6 January 2021 US Capitol riots, tends to explore individual-level impact rather than the kinds of collective structures ¹ that an approach like Mernissi's is well-suited to address. The atomistic Western model of social contagion, as opposed to communal debate, may be traced at least partially to the deeply problematic ideals of social scientific neutrality and objectivity espoused by the Enlightenment—ideals that have been uncritically adopted by institutionalized data science discourse (Crawford, 2021).

The fragile but well-guarded self-image of data science as devoid of ideology allows it to perpetuate the illusion of value-free objectivity and present its “scientific” output as fact. In the context of algorithmic ideology, or the self-fulfilling hypothesis-testing process that enables researchers to produce the results they set out to find proof for, practically any belief can be turned into a truth about reality with the right amount of computational calibration (Mager, 2012). This behavior is certainly not new and can be traced back to Enlightenment thinking, which supplies us with a universalized subject of knowledge, an empirical orientation to phenomena, and a belief in the universality of reason, leaving no room for alternatives (Castle, 2009). What constitutes a new layer to this ostensibly impeccable formula of objectivity is the power of computation, which, with its “black box” models, affords data scientists plausible deniability to say, should anything go wrong, that it was the algorithms’ fault (Pasquale, 2015). CD&AS scholarship problematizes this logic by exposing the human subjectivity of algorithms. It points, for instance, to the “warm human and institutional choices that lie behind these cold [technological] mechanisms” (Gillespie, 2014), thus advancing the understanding of algorithms as socially determined. Transnational feminism allows us to continue this critical discourse by further contextualizing the development of data science not as an objective temporal process but rather as a long historical development which involves the selection and exclusion of ideas, figures, and interests by cultural elites (editors, academics, corporate officials, and grant-awarding institutions) as well as the tendency to develop canons according to the shifting criteria of the “marketplace of ideas,” institutional memory, and social and political power (Maddox and Malson, 2020).

• 4.

Diversifying Knowledge

The university is where the power of cultural elites is most palpable nowadays (Castle, 2009); although Silicon Valley glorifies the figure of the anti-academic college-dropout such as Bill Gates or Mark Zuckerberg, university placement records tell a different story. It is evident that many occupants of Palo Alto's tech desks have received at least partial computer science education, and it can therefore be assumed that they have been exposed to the same principles of computation as a masculinist, God-like activity that continue to dominate Western computer science campuses (Crawford, 2021).

In the context of decades-long algorithmic colonization of people's most intimate data, diversifying the data science workforce, as D’Ignazio and Klein (2020) propose, may not be sufficient. What needs to accompany this process is the diversification of knowledge itself, which can help to radically rethink the very foundation of computation (Moats and Seaver, 2019) and oppose the forces of data colonialism (Couldry and Mejias, 2019). Transnational feminism provides us with analytical tools to critically examine the reality that data science envisions and the origins of this ideological framework. It allows us, for instance, to recognize something which Data Feminism does not go far enough to uncover, namely, the possible misalignment of teleological incentives for data scientists on the one hand, and anyone claiming a feminist understanding of the world on the other. The thinking and actions of the former are based on maximizing accuracy, whereas for the critical feminist subject, the goal of any type of work, data or otherwise, is maximizing justice (Mohanty, 2003). These radically different orientations often collide because they focus on different objectives—developing a more productive society versus a more equitable one. Moreover, the clash between these two systems of thinking is inevitable, since algorithms allow the logic of the neoliberal marketplace to infiltrate all areas of life, such that we now live in a world where the same logic Google uses to rank webpages is leveraged to single out potential repeat offenders by justice systems across multiple states in the United States (Washington, 2018).

• 5.

Demystifying the “Newness” of Algorithms

In this sense, algorithmic domination is more of a regress to naive empiricism than a radically new and innovative way of decision-making, and it is therefore representative of same-old Enlightenment thinking more so than any kind of postmodernity (Crawford, 2021). It is then perhaps best described not as a paradigm shift, but a different delivery system of oppression marked by invisibility and inscrutability (Browne, 2015). Data Feminism has gathered as much, given its critique of Enlightenment logics. However, it falls short of calling to task two major aspects of Enlightenment thinking which continue to fuel the data “revolution” today—capitalism and settler colonialism (Ali, 2016). It can thus offer only epidermic interventions into the dominant data science regime: always within the framework of tools offered by it in the first place.

While the perniciousness of algorithmic decision-making has been interrogated in data feminism, a transnational feminist perspective adds to this debate a supra-Western analysis of algorithmic categorization—separating data points (too often—people) into distinct classes. The objective of one of the most widely used classification algorithms—support vector machines—for instance, is to find “the best separating line” for data, such that the distance between the points on each side of the line is maximized. In cases where such a line cannot easily be identified, the data is transformed into a higher-dimensional space (e.g. from two to three dimensions) in which case the separation zone is demarcated by a three-dimensional “optimal separating hyperplane” instead of a two-dimensional line (Hofmann, 2006). It should come as no surprise that historically marginalized persons often end up on the wrong side of the “hyperplane,” but also that such algorithms—trained on Western-centric datasets—get deployed in the Global South where they perform dismally and produce harm (Sambasivan et al., 2021). Thus, if the problem of the 20th century was the color line (DuBois, 1986), one could say the problem of the 21st century is the optimal separating hyperplane, which in a vortex-like fashion embodies the color line, but also the “lines” of gender, sexuality, class, ability, religion, ethnicity, and nation-state.

Conclusion

While (neo-)colonial powers continually mark territorial national borders, a much more insidious yet no less powerful process of (re)mapping occurs in data science. Why would data scientists not consider themselves gods incarnate, when they, with a single model calibration, get to make life-changing decisions such as who is marked for a jail sentence? A transnational feminist coalitional ethos provides insights and opportunities to think about alternatives to the current system and bring awareness to the necessarily political nature of data science. Among other contributions, it demonstrates that a passive, institutionalized, depoliticized version of data feminism can perhaps attract followers who do not want to bother with collective organizing and prefer a tamed belief system “from the comfort of the armchair,” but will certainly not galvanize the kind of global movement necessary to resist the ubiquitous effects of algorithmic oppression. Data feminists ought to strive for a truly transnational data science framework.