Cultural Gradients and the Sociotechnics of Data

The advent of what business strategists are fond of calling the “Cambrian explosion” in artificial intelligence (AI) has spawned a legion of concerns about downstream effects. It’s hardly an argument at this point to say AI can have unintended, even harmful consequences for social life. The many calls to make AI fair and accountable, transparent and ethical, attest to such impacts, or the threat thereof. Acting on these calls often requires description and explanation: what, upstream, caused downstream effects? But as Jonathan Roberge and Michael Castelle claim in the introduction to their edited volume, The Cultural Life of Machine Learning, the up-/downstream dynamic, when used as an explanatory frame, tends to uphold what is “largely (if unconsciously) a positivist project” (12), one for which the harms of AI ultimately require technical solutions. It does so, they explain, by eliding AI’s own “sociotechnical genesis” (3), the dense mesh of decisions, practices, and technologies that entangle AI in social factors from the start. This elision makes the (upstream) technical seemingly detachable from the (downstream) social; as a “corrective,” Roberge and Castelle propose “what could be called—with a wink and a nod to deep learning methodology—an end-to-end sociology of contemporary AI/ML [machine learning]” (3). The chapters in their volume accordingly bring this proposal into being.To clear the ground for this end-to-end approach, Roberge and Castelle provide a series of compelling pairings between the conceptual inventory of AI/ML and that of sociology, media studies, and science and technology studies. If, as they write, ML practitioners “see their own behavior in terms of the epistemology of their techniques” (6), a sociological critique of AI/ML would begin by assessing the overlaps and gaps between adversarial deep learning and deep play (Clifford Geertz), concept nets and normative classes (Geoffrey Bowker and Susan Leigh Star), explainable AI and hermeneutics (Hans-Georg Gadamer), algorithmic nudging and actor networks (Bruno Latour), and learning bias and social prejudice (Joy Buolamwini and Timnit Gebru). Literature scholars may find a last pairing to be especially productive: “co-text” and context, the raw co-occurrence of “entextualized” symbols and the meaning of that emplacement (8). Indeed, for Roberge and Castelle, determining what and how context means is where AI/ML and their own approach are most linearly separable.Several of the collection’s chapters offer rich historical accounts, emphasizing tenuous formations of epistemological consensus among ML practitioners. Aaron Mendon-Plasek shows how character recognition was key to the development of “contextual significance” in early AI, because it bolstered a strategy of explaining away systems’ deficiencies with highly accurate results for constrained tasks (39). Théo Lepage-Richer locates an abiding concern with managing uncertainty in neural models and Cold War cybernetics; this, he claims, paved the way for a wider “adversarial epistemology” that recasts failure as an “opportunity” to improve models (218). In contrast, Ceyda Yolgörmez draws from Alan Turing and George Mead to theorize how we “encounter” machines in a “social sense” at the very moments cyberneticians sought to avoid, where calculative processes disrupt expectation (150). Other chapters discuss the politics of learning behavior. Fenwick McKelvey tracks the rise of a “new political science,” primarily with regard to the Simulmatics Corporation, one of the first consulting groups to use computer simulation as a “proxy for public opinion” (119). And in a deep engagement with the educational theory of Lev Vygotsky, Tyler Reigeluth and Michael Castelle probe the core ideas of concept development and generalization in ML. Vygotsky’s own elaboration of these ideas in the 1920s and 1930s positions Reigeluth and Castelle to identify a missing piece in how machines learn: a necessarily social element that “pulls” learners into a “zone of proximal development” (95).A last set of chapters turn to the present. Werner Binder highlights the braided agencies attending AlphaGo’s entry into professional Go. While recounting how the system won against the world’s top Go players, Binder unravels key ambivalences in players’ reactions to the games, including their adoption of AlphaGo’s more novel play strategies. Using place-based research conducted in Chile’s Atacama Desert, Orit Halpern returns to uncertainty, suggesting that the many networks “bridg[ing] data and matter” in planetary-scale AI (from the Event Horizon Telescope to speculative resource extraction) have triggered an epistemic break in contemporary governance (228). Luke Stark, Daniel Greene, and Anna Lauren Hoffmann close the volume by surveying how state and corporate governance mechanisms for AI/ML currently forestall wider critique. The authors conclude that, in place of technical fixes or values statements, such mechanisms must begin by “foregrounding and centering the expertise” of communities directly affected by AI/ML (272); the Tech Won’t Build It campaign and the Movement for Black Lives exemplify such an approach.Amid the extensive range of these essays, one black box does go mostly unopened: data wrangling. Save for a handful of brief references, the collection skips over the vast amounts of time and energy involved in preparing data for models. In view of this, Data / Set / Match, an exhibition program held at the Photographers’ Gallery in London from 2019 to 2020, adds an important dimension to Roberge and Castelle. Introducing the program, Nicolas Malevé (2019) writes, “The automation of vision [by ML] has not reduced but increased the number of eyeballs looking at images, of hands typing descriptions, of taggers and annotators.” The essays, talks, and commissions that comprise Data / Set / Match recognize this increase as a core feature of contemporary visual culture. Together, explains Malevé, they explore how the work of “collecting, labelling, composing, assembling,” and finally “distributing” images—all done to create “ground truth” data for models—might be traceable in giant data sets like ImageNet or COCO (Common Objects in Context).A shared strategy among the exhibiting artists is superimposition. In Tunnel Vision, Philipp Schmitt uses custom software to identify objects in busy 35-mm scenes and overlays them with disparate matches from COCO. Accompanying the latter are a label and a username, reminders that ML is vestigial human interpretation. Part performance, part curation, Everest Pipkin’s Lacework samples from Moments in Time, a set of 1 million three-second videos depicting actions associated with verbs. Pipkin watched them all; Lacework presents an algorithmically “upscaled” selection, which adds a blurring effect to simulate, perhaps, the tired eyes of digital piecework or, as with the abstract visages of Heather Dewey-Hagborg’s How Do You See Me?, a computer learning to see. The “speculative remixes” of quotes from Virginia Woolf and Katherine Anne Porter, which xtine burrough and Sabrina Starnaman set atop clips from the Epic Kitchens data set in A Kitchen of One’s Own, tether data annotation to a longer history of feminized, unpaid labor. But it is Mimi Ọnụọha who presents digital piecework in the starkest terms: The Future Is Here! shows images of the domestic spaces where “sub-sub-contractors” annotate data (Schmidt 2020). Periodically, the images flip into the style of a comic book, complete with canned slogans. “A world unlike any other!” Yet, to be sure, Ọnụọha’s inking and neon palette also recall the encroaching outlines of modern image segmentation algorithms.On display throughout Data / Set / Match are the two streams, up and down, upon which Roberge and Castelle’s collection is also trained. The exhibition’s many facets collectively hold these streams in tension, offering a way to “learn the rules and edges” of a data set (Pipkin’s Lacework), to catch a glimpse at the furtive but pervasive human life bound up with AI/ML. In this sense, it performs considerable, if tacit, sociological work about the sociotechnics of data. For those who may be grappling with how to bring this or the many other social entanglements of AI/ML into view, both Data / Set / Match and The Cultural Life of Machine Learning provide—and themselves serve as—vital context for future inquiry

Geography · Ethics and Social Impacts of AI · History