Sunday, September 4, 2022

Saving Face: Portraiture in the Age of Artificial Intelligence

THE PHOTOGRAPHER approached the trio on his bike, a bundle of gear balanced on his front fender and a wide-brimmed hat covering his head. A discussion ensued. Hubert, the first of the three to speak, took the man for a fellow enthusiast of far-left politics. Peter suspected a questionable sales pitch. Adolphe, at the head of the cluster, offered apologies for his companions’ brusqueness. No worries, the photographer countered; he was there, in fact, to propose a group photo—a slice-of-life shot that, he insisted, would be “for science’s sake only, and for the archives: a personal record of the conversation we shared today.”1

Such is the encounter of German photographer August Sander and the three young men pictured in one of his best-known images—Jungbauern (Young Farmers), 1914—as imagined in the opening pages of Richard Powers’s 1985 debut novel, Three Farmers on Their Way to a Dance. As Powers’s title suggests, Sander’s image functions as both the spark for and the through line of his novel, which begins with a viewing of the photograph decades later at the Detroit Institute of Arts and subsequently traces interwoven narratives that follow Sander’s farmers across time and space: to the country dance they were headed for on that 1914 spring day, to the cataclysmic war that began only months later, and finally—through both their lived actions and their photographic representation—to the experiences of later generations on both sides of the Atlantic.

Our Instagram feeds and work-profile pages, in short—where each newly uploaded photograph marks a potential addition to expanding training data sets—are the Texas oil wells of the Clearview AI age.

Young Farmers, most readers will know, is among the hundreds of photographs, made across decades, that Sander came to conceptualize as part of his monumental (and never realized) portrait project “People of the Twentieth Century.” Conceived around 1922 as a set of forty-five portfolios of twelve photographs each—to be arranged in a carefully assembled sequence moving from farmers to craftspeople and the educated middle class to, finally, a diverse array of social outcasts labeled simply “the last people”—Sander’s vast survey was to present a systematic, if not quite comprehensive, sociological atlas of the German people. Or, more precisely, of the German people (who had lost their kaiser only with the war’s end in 1918) on the path to democracy, advanced industrial modernity, and, ultimately, authoritarian terror.

It was just this aspect—his portraits as the index of a specific culture in the throes of a particular historical process—that Sander chose to highlight in titling his preliminary 1929 sampling of his portraits with the expression Antlitz der Zeit: roughly translated, “the face of our time.” This dimension of the work was at the root of critic Walter Benjamin’s interest in Sander’s project as well. In his seminal essay “Little History of Photography,” published two years later, Benjamin praised Sander’s photography as providing nothing less than a quasi-physiognomic “training manual” for the age—an era, the critic famously wrote, in which individuals “will have to get used to being seen according to where one comes from.”2 Sander’s photographs of bakers and accountants and washerwomen, in other words, pictured for Benjamin less isolated subjects than a distinctly modern logic of subjectivity itself—in which what we are and how we come to be seen is indivisible from the cultural environments, markers, and relations that form and connect us. To return to Powers: Those young farmers, from their dangling cigarettes and just-so stares to their pinched fingers and matching suits, invite us—as do their dozens of compatriots in Sander’s atlas—into an interconnected world of spring dances and August military campaigns and, should we accept the novelist’s wager, into the imagined afterlives of these figures and events across cultures and continents in the decades to follow. (This past August, Aperture published the most comprehensive iteration of Sander’s unfinished project to date, including 619 of his photographs in a single volume.)

THE USE OF VAST PORTRAIT COLLECTIONS as essential training manuals is, today, a much more pressing and widespread phenomenon than Benjamin could have ever imagined. To take just one notorious example from recent headlines: The facial-recognition company Clearview AI, launched with assistance from alt-right bottom-feeders including Peter Thiel and Charles Johnson, recently notified investors that it aims to amass one hundred billion facial photographs for its database by the end of the year.3 These images, culled without consent from across the internet, serve as training data sets for machine-learning systems whose goal is to be able to classify and identify any face on earth, allowing for precision-guided and near-total surveillance. The Ukrainian government, for instance, has recently used Clearview AI technology to identify Russian war dead, just as grocery magnate John Catsimatidis in 2018 used it to spy on his daughter’s date. Other Clearview AI clients include the NBA, Macy’s, and—of course—various federal, state, and local law enforcement agencies. In China, a similar project is taking shape under the guidance of the country’s Ministry of Public Security, which is integrating footage from more than half a billion surveillance cameras with an unprecedented sampling of iris scans, voice prints, and DNA samples—all, in the authorities’ own words, with the goal of successfully “controlling and managing people.”4

The billions of images collected by Clearview AI and the Chinese government (and other, hopefully more benevolent machine-vision devotees) function to construct what artificial-intelligence developers refer to as a “ground truth”—a set of verifiable data that establishes the epistemic boundaries of machine-learning processes and facilitates the construction of algorithms by which inputs (selected images) are matched with outputs (targeted classifications and identifications). These outputs range from straight identification (this picture shows that person) to predictive judgments (this picture shows somebody reliable, likely criminal, etc.). To continue to fine-tune the algorithms reaching these conclusions, of course, one needs more and more data—hence the relentless and unregulated drive to cull billions upon billions of images, from publicly available mug shots and security footage to Instagram selfies. (Indeed, everyone reading these words is likely included in Clearview AI’s data set.) Even more, this process necessitates the constant creation of data—the “(total) datafication and surveillance,” as Jathan Sadowski has written, “of people, places, processes, things, and relationships among them.”5 Our Instagram feeds and work-profile pages, in short—where each newly uploaded photograph marks a potential addition to expanding training data sets—are the Texas oil wells of the Clearview AI age.

As AI researcher Kate Crawford has noted, the metaphor of data as natural resource can be problematically bloodless, diverting our attention from the very real material and human costs of data’s extractive economy. Crawford (who with Trevor Paglen organized the 2019–20 exhibition “Training Humans,” focused on facial-recognition training data sets, at Milan’s Fondazione Prada) has detailed the extent of these costs, which reach from the geopolitical violence generated by the natural-resource mining essential for AI infrastructure to the ethical vacuum resulting from the replacement of the notion of human subjects by that of data subjects—yielding a vision of individuality, in Crawford’s words, that understands only “agglomerations of data points without subjectivity or context or clearly defined rights.”6 With such an understanding, it follows that “the meaning or care that might be given to the image of an individual person, or the context behind a scene, is presumed to be erased at the moment it becomes part of an aggregate mass that will drive a broader system.”7

This schema is the exact opposite, of course, of that encapsulated in Sander’s three farmers as conceptualized by Benjamin and narrativized by Powers—for whom, just as for the photographer himself, it was the complexities of subjectivity and context unlocked by the 1914 photograph that made it so compelling, and which, when compiled in a larger “training set” with such images as Sander’s Putzfrau (Cleaning Woman), 1928, or Ingenieur und Werbeleiter (Engineer and Advertising Manager), ca. 1935, raised fundamental questions about the economic, political, and social structures by which rights, in a country that had only fully embraced democracy with World War I’s end, might be established and maintained at all. Note the stare, garb, setting, and held implements of each of Sander’s represented subjects. They are utterly distinct and yet strikingly parallel; each individual or group both opens up a specific life-world to which they belong (one into which Sander himself entered, and which is ripe for a novelist’s gaze) and establishes dialogic correspondences with the others—creating an active portrait of the society they, and Sander, composed and inhabited together and of the commonalities and contestations this cohabitation necessarily entailed.

To face this landscape, a new program of physiognomic analysis and instruction is in order: one that examines both the countenance of AI itself and the amalgamated complexity of the individuals and communities it seeks to reduce to algorithm-building data.

As Benjamin helps us to understand, the insistent play of material specifics across Sander’s faces operates directly counter to another technologically mediated facial economy from Germany’s interwar years—that of the fascist face. And this distinction helps, in turn, to enlighten the political stakes of our own AI-driven age. Just as today’s rapacious collection of data is fueled by the need to dissolve any and all complexity of subjects, contexts, and relations within the unifying system of predictive algorithms, so the fascist economy of the face that characterized Hitler’s Germany operated through a logic of the many dissolved into the one. The individuated faces of the country’s farmers, cleaning ladies, and businessmen, that is, came to be phantasmatically reflected and sublimated in Hitler’s own visage—a symbolic structure exemplified nowhere so clearly as in the call-and-response staged between the assembled laborers and the German führer in Leni Riefenstahl’s 1935 party-rally documentary, Triumph of the Will. In the integrated essence of the Volk that Riefenstahl’s film aspires to capture, there is no room for individual detail or imaginative extension; the system it emblematizes is both anchored and sustained by a total identification of mass and leader. “Where would the Volk be,” as Bertolt Brecht imagined Hitler wondering to his mirror reflection, “without the luck of having me reflected in it, and it being able to be reflected in me?”8

This identificatory apparatus, as Benjamin among others argued, worked by taking the place of—and violently suppressing—the messy back-and-forth of politics. No wonder Sander’s studio was raided by the Gestapo and the plates for Antlitz der Zeit seized in 1936; does anything betray the fallacy of the integrated Volk more effectively than our conversant trio of businessman, washerwoman, and farmers? In today’s promises of an integrated mirroring of the world in AI’s assembled data, we see an analogous logic of absolute, and hence depoliticized, identification taking shape—one that operates through both individual subjects’ identification with AI’s claims to unfettered, all-knowing ability and, simultaneously, the ambition to identify (and in part control) these subjects via AI’s ever more expansive surveillance capacities. As Crawford astutely outlines, these dual identifications point to the double-sidedness of what she and the historian Alex Campolo have called the “enchanted determinism” of AI: the fact that AI systems, in their mix of infinite aspiration and mathematically driven systematicity, are seen to be at once “enchanted, beyond the known world, yet deterministic in that they discover patterns that can be applied with predictive certainty to everyday life.”9

A primary problem with such thinking, just as with the closed mimetic loop showcased by Riefenstahl, is that it obscures the fundamentally political decisions, priorities, and prejudices in which such enchantment is anchored and to which it gives rise. For all the benign and even beneficent uses of AI-assisted machine vision (from self-driving vehicles to medical imaging tools), to downplay the fundamental political stakes of its history, capacity, and goals is to make fatally easy the slip into totalizing surveillance as a primary function. Is it any surprise that the cast of characters filling facial-recognition technology’s brief history often reads like a rogues’ gallery of far-right ideologues, from Johnson and Thiel to Trump megabacker Robert Mercer? Or that the algorithms guiding this technology have been notoriously inaccurate, or simply incapable, when confronted with nonwhite faces? Or that the very process of training these algorithms relies on a binary logic of “this, not that” that cannot admit any porosity in the categories through which faces—and individuals—might be understood? As Florian Jaton titled his 2017 article on the design of image-processing software, “We get the algorithms of our ground truths.”10 Exactly. And these ground truths, we can be sure, are permeated by the very same prejudices, priorities, and power differentials from which, we are told, AI’s wonders promise a blissful escape.

These developments, it barely need be said, mimic—and help drive the larger logic of—neoliberal rationality’s economization of the political and (in Wendy Brown’s words) “jettisoning of the very idea of the social.”11 Rather than engaged citizens, this structure values reliable “team members” for whom the active contestation of rights, responsibilities, and values (the old business of citizenship) is superseded by an overarching concern with the economic growth of the nation-as-firm and with the securing of their roles in this process (the new business of human capital). Mirrored and emblematized in the enchanted determinism of AI—a primary new function of which is tracking the precise actions of such team members, at Amazon and elsewhere, to ensure their efficient servicing of the system to which they’re subordinated—this unifying economistic framework shunts, and works to supplant, the component differences of which a citizenry is made.

To face this landscape, a new program of physiognomic analysis and instruction is in order: one that examines both the countenance of AI itself and the amalgamated complexity of the individuals and communities it seeks to reduce to algorithm-building data. We can see attempts at the former in such recent projects as Shalini Kantayya’s 2020 film Coded Bias and Crawford and Paglen’s “Training Humans” show—both of which assiduously analyze the creation and operation of AI algorithms to, in the words of the latter’s website introduction, “look back at these systems, and understand, in a more forensic way, how they see and categorize us.”12 Beyond forensic analysis of AI’s epistemic boundaries and flaws, we need renewed epistemologies of community itself—ones that aim not to distill complex realities into simplified formulas but, like the work of Sander, to build complexities of relation and recognition from the seemingly simplest of means: the human portrait.

McQueen’s Year 3 photographs, shot in the wake of Brexit and installed right before Covid shut everything down, presented an overwhelming view of subjects and community in formation.

Outstanding examples of such work have, perhaps unsurprisingly, flourished in the art of the past three decades, precisely those years in which the internet—and the model of data subjectivity it promulgated and continues to fuel—has come to dominate our lives. These range from Rineke Dijkstra’s “Beach Portraits,” 1992–2002, to diverse series by Dawoud Bey, LaToya Ruby Frazier, Lyle Ashton Harris, Catherine Opie, Thomas Struth, and others, to Steve McQueen’s astounding 2019–21 project Year 3, in which the British artist filled the Tate Britain’s monumental Duveen Galleries with simple group portraits of third-grade classes—public, private, parochial—from across London. (The goal was for every such class in the city to take part; around two-thirds did.) An immense sea of postures, getups, body types, and classroom dynamics, McQueen’s assembled photographs, shot in the wake of Brexit and installed right before Covid shut everything down, presented an overwhelming view of subjects and community in formation at a time of all-consuming social stress: a profoundly moving assemblage of utterly diverse individuals, right on the cusp of adolescence, united by their shared grade and city. The essential “ground truth” his project presented was that of the impossible irreducibility, and forever coming-into-being, of human subjects and groupings. That McQueen achieved this with the most familiar of institutional formats—the class portrait, a not-too-distant cousin of the mug shot—made his project all the more effective. Across picture after picture, this rigid framework amplified its own undoing in the profusion of incident and idiosyncrasy that filled it—and in the process made clear the distinction of such historical portrait models from AI-guided machine vision, in which the foundational act of making-into-data seeks to eliminate, precisely, any disturbance from such idiosyncratic “noise.”

Of course, there’s a decent chance that many of McQueen’s Year 3 shots are now doing their part in Clearview AI’s training data set, as may the individual portraits of Bey, Dijkstra, Frazier, Harris, Opie, and others. That such “making data” may have occurred, though, doesn’t mean it inevitably must. Rather, the opposition of McQueen’s third graders as specific people and potential data-set training material clarifies the stakes: We must prioritize—and call on state power to protect—individual subjects over algorithm-driven systems, and the complexities of social difference over the unifying drive of abstract capital. We can’t, for all Powers’s novelistic imagination, ask Sander’s three farmers for their thoughts. But we can, entering their indelible image and the history of which it is a part, begin to map the terrain we ourselves are crossing and chart the course we wish to take.

Graham Bader is professor of art history at Rice University in Houston. His book Poisoned Abstraction: Kurt Schwitters Between Revolution and Exile was published by Yale University Press last fall.

NOTES

1. Richard Powers, Three Farmers on Their Way to a Dance (New York: William Morrow and Company, 1985), 27.

2. Walter Benjamin, “Kleine Geschichte der Photographie,” in Gesammelte Schriften, ed. Rolf Tiedemann and Hermann Schweppenhaüser, vol. 2, pt. 1, Aufsätze Essays Vorträge (Frankfurt: Suhrkamp, 1977), 381. Translation my own.

3. For details, see for instance Drew Harwell, “Facial Recognition Firm Clearview AI Tells Investors It’s Seeking Massive Expansion Beyond Law Enforcement,” Washington Post, February 16, 2022, washingtonpost.com/technology/2022/02/16/clearview-expansion-facial-recognition/.

4. See Isabelle Qian, Muyi Xiao, Paul Mozur, and Alexander Cardia, “Four Takeaways from a Times Investigation into China’s Expanding Surveillance State,” New York Times, June 21, 2022, nytimes.com/2022/06/21/world/asia/china-surveillance-investigation.html.

5. Jathan Sadowski, “When Data Is Capital: Datafication, Accumulation, and Extraction,” Big Data & Society, 6, no. 1 (2019): 2.

6. Kate Crawford, Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence (New Haven: Yale University Press, 2021), 115.

7. Ibid., 93.

8. Bertolt Brecht, “Notizen zu Heinrich Manns ‘Mut,’” as cited in Gerhard Richter, “Face-off,” in Monatshefte, 90, no. 4 (Winter 1998): 418. Translation my own. Richter’s essay is a highly intelligent discussion of the logic of the face in German culture and criticism of the Weimar and Nazi eras.

9. Crawford, Atlas of AI, 213.

10. See Florian Jaton, “We Get the Algorithms of Our Ground Truths: Designing Referential Databases in Digital Image Processing,” in Social Studies of Science,<span class=“Apple-converted-space”>  </span>47, no. 6 (December 2017), 811–840.

11. Wendy Brown, Undoing the Demos: Neoliberalism’s Stealth Revolution (New York: Zone, 2015), 210.

12. On Crawford and Paglen’s show, “Kate Crawford | Trevor Paglen: Training Humans,” Fondazione Prada, fondazioneprada.org/project/training-humans/?lang=en.



from Hacker News https://ift.tt/a8wBLmb

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.