Chun 2021
Chun, Wendy. Discriminating Data. 2021
Introduction : how to destroy the world, one solution at a time
Hopeful ignorance is not the s tion but the problem: it perpetuates discrimination and inequality, one solution at a time. The problem is not that giant technology monopolies have disrupted habits, institutions, and norms in order to create new, unforeseen futures. The problem is that, in the name of “creative disrup- tion,” they are amplifying and automating—rather than acknowledging and repairing—the mistakes of a discriminatory…
Page 2
The basic story line was this: naive hippies fall in love with libertarians, hook up with Wall Street sharks, and inadvertently destroy the world in their attempt to keep it free.
Page 5
In post–World War III Neuromancer, inequality and violence predominate; a criminal underclass has replaced the working class; and the United States is no longer a nation-state. So how did this apocalyptic vision—written in the shadows of the Cold War, the coming nuclear annihilation, and the “Japanification” of the world— become utopian? What made it so attractive to those who would become “the Internet”?
Page 7
Written to coincide with the Davos Forum and “24 Hours in Cyber- space,” a 1996 media event, John Perry Barlow’s “Declaration of the Inde- pendence of Cyberspace” is perhaps the most iconic description of the Internet reborn as cyberspace.
Page 8
This declaration of independence conceptually transformed a military- educational network, built by the U.S. government, into a bodiless—thus “privilege free”—space of freedom, escape, and libertarian self-interest. It also portrayed Silicon Valley elites as militan…
Page 9
By becoming cyberspace, the Internet became an “electronic fron- tier” and thus a wilderness ripe for settler colonialism and exploitation, and, as Jodi Byrd has argued, for the reemergence of “natives” without natives.
Page 9
This rhetoric may seem dated, yet its power and hopeful ignorance remain and make themselves felt in statements that conflate empower- ment with bodily escape, and it drives an endless game of hide-and-seek, rebellion, and punishment. 34It misidentifies Silicon Valley acolytes as rebels or underdogs, regardless of their actual circumstances or obscene …
Page 10
Do we really want Silicon Valley to be responsible for our future? What else will it take in the name of accountability?
Page 11
Most succinctly: escape for the few and misery for the majority are goals, not unfortunate errors.
Page 12
To dispel this “sovereign” nightmare, we need to understand how the desire to erase race and difference perpetuates discrimination and inequality. We need to comprehend how histories of slavery and inequal- ity fuel the nightmare of supreme sovereignty and the opposite side of its coin: AI as the coming apocalypse in which masters become …
Page 12
According to many scientists, technologists, and science fiction writ- ers, “AI=The Apocalypse.” It ends human work; it ends human freedom; indeed, it ends everything human.
Page 12
44With each revolution, well-paid or relatively well-paid jobs in the global North have become less well-paid ones elsewhere, from programming to data entry to circuit building.
Page 13
It is no accident that those developing and intimately intertwined with technology were, and are, both the most fearful and the most certain.
Page 14
Coined by Karel Capek in his 1920 play R.U.R, “robot” comes from robota, the Czech word for “forced labor.” Written during the time of Communist ferment, Capek’s play centers around a rebellion, in which the victorious robots declare: “Robots of the world! The era of man is at an end! . . . A new era has begun! . . . Salute Robot rule!” 52As literary critic Jenny Rhee has argued in The Robotic Imaginary: The Human and the Price of Dehumanized Labor, the enduring power of raced and gendered robots within the cultural imagination, as well as within science, technology, and engineering, is linked to the history of slavery.
Page 14
The history of slavery is central to the history of computing. Con- trol systems were first called “servo-mechanisms.” “Master” and “slave” functions and circuits riddle computers.
Page 15
World-destroying liberation envy, however, is not the only solution. Engaging Indigenous knowledge and histories would place current crises within the larger context of colonial expansion. 62Notions of dystopian destruction and surviving the apocalypse are not new; rather, they stem from the very emptying of Indigenous lands into the “New World”—a move that haunts “new media” and its frontier dreams. 63By following rather than usurping struggles for equality and freedom, we can move from apology to reparations, from dreams of escape to modes of inha…
Page 16
It is because our current society is so unequal that it seems easier to imagine the end of humanity than the end of injustice or capitalism.
Page 16
To inhabit this world together, we need—among so many other things—to understand how machine learning and other algorithms have been embedded with human prejudice and discrimination, not sim- ply at the level of data, but also at the levels of procedure, prediction, and logic, one apology at a ti…
Page 16
As Kate Crawford and legal scholar Jason Schultz have shown, big data compromises the privacy protections afforded by the U.S. legal system by making personally identifiable information about protected categories, such as gender and race, legible.
Page 17
Given these programs and U.S. legal protections, many analyses have focused on revealing proxies that implicitly index race in explicitly color-blind systems. As these examples and work by sociologists such as Eduardo Bonilla-Silva on color-blind racism have shown, “ignoring” explicit markers of race amplifies—rather than alleviates—racism. 84Not only does it lead to a sit- uation in which racism is naturalized; it also embeds whiteness as defau…
Page 20
A clear example of this is facial recognition technology (FRT), which has been repeatedly—and justifiably—accused of racism for its recognition defects (see chapter 4).
Page 20
The p lem stems from the libraries on which these algorithms have been tradi- tionally trained: the “ground truth” for these programs are the faces of Hollywood celebrities and university undergraduates, those well-known hotspots of diversity (figure 4). At a fundamental level, this “curation” means that ground truth = …
Page 20
The question is not why is this happening? but rather why is this still happening?
Page 22
These “errors” come from “ignoring” race—that is, by assuming that race-free equals racism-free. The solution, however, is not simply the explicit inclusion of race within these programs—programs that better recognize black faces will not solve the problem of discriminatory polic- ing. So, how do we fight racism and its proxy …
Page 22
Discriminating Data responds to this question by interrogating assump- tions embedded within network science and machine learning as they are currently configured regarding segregation, discrimination, and history.
Page 22
My goal throughout Discriminating Data is to help release us from the seeming vise grip of preemptive futures by using critical theory, statistics, and machine learning tools probingly and creatively. Rather than con- demn these tools as inherently eugenicist, I seek to understand the tools’ limitations and possibilities by engaging their log…
Page 25
It explains that the move from “mass media” or mass society, marked by ambivalence and neutrality, to polarized networks, marked by angry resistant clusters, is fundamental to the history and design of social networks.
Page 27
Red pill toxicity, or, Liberation envy
As researchers across the many disciplines working on mis/disinformation have noted, fact checking, though important, is sim- ply not enough. Fact-checking sites are undermined by the structure and speed of global communications: they lag behind the deluge of rumors produced by networked disinformation sources and spread through pri- vate interac…
32
Rather than simply dismiss this twinning of misinformation and authenticity as irrational or accidental—a roadblock to analysis—what if we took the twinning to be formative and general?
32
” Questioning information is not antidemocratic—critical thinking grounds democratic education. The danger is not mistrust or criticism, but rather the transformation of mistrust into a deep faith in dubious sources.
33
This slavery theme, resonating with the enslaved white men in Apple’s “1984” commercial, conflates the human race with the black race: it transforms the hacker network into an “underground railroad,” an invisible system of resistance struggling against an invisible world of power. The white Agent Smith makes this parallel absolutely clear during his “talk” with Morpheus, which repeats racist statements of smell and taste.
33
Cyberspace, as a reparative hallucination within a hallucination, enables dominant groups to dis-identify as oppressed, mili- tant minorities through hopeful ignorance. What was once “majority culture” or “mainstream” has become fractured into agitated subcultures that nonetheless cohere into an angry dominant ideology. Power can now operate through reverse hegemony: if hegemony once meant the creation of a majority by various minorities accepting a dominant worldview (such as most of the Greek city-states accepting Athenian values), 17now hege- monic majorities can emerge when angry minorities, clustered around a shared stigma, are strung together through their mutual opposition to so- called mainstream …
34
18The point is never to be a “normie” even as you form a norm.
34
Correlating eugenics
Big data’s power was said to be based on correlation, but this was not correlation’s first rodeo. Along with linear regression and other founda- tional statistical methods, correlation was developed by early twentieth- century biometric eugenicists, who were eager to breed a better “human crop.” By investigating the historical ties between big data and eugenics, we will see that the two are linked together by a fundamentally undisrup- tive view of the future. But, as we will also see later in the chapter, even though both have sought to make the future repeat a highly selective and discriminatory past through correlation (so that ground truth = deep fake), they differ in several important respects. In the transition from eugenics to data analytics, the focus group moved the nation to the neighbor- hood/tribe; the goal shifted from uplift to escape; and homophily (the notion that similarity breeds connection) went fro…
36
… Psychographics superseded demographics, geographics, and economics, with their crude assumptions “that all women should receive the same message because of their gender, or all African Americans because of their race, or all old people, or rich people or young people to get the same message because of their demographics.
36
The goal of the firm’s advertisements was to create transformational, “red pill” experiences: to have users go “down the rabbit hole” by fol- lowing ads, carefully “breadcrumbed” across different sites and spread by their friends and others “like them.” 12Identification—or targeting—…
40
the first function within the program: the others were recognition (mutual identification) and conversion.
41
If we have learned anything from Cambridge Analytica, however, it is that finding and exploiting clusters is key to cultivating individual behav- ior—to fashioning change. “Personalization” works at the levels of individ- ual actions, “latent” factors, and “bespoke” network neighborh…
41
P dictive data analytics for Internet users work—if and when they do—not by treating every Internet user like a unique snowflake, but rather by seg- regating users into “neighborhoods” or petri dishes based on their slightly odd or deviant—that is, “authentic”—likes and dislikes. Individuals are formed and identified by their so-called neighbors. 14Cambridge Analytica claimed to have discovered proxies that revealed a person’s race, sexual orientation, political leanings, and so on: preferring an American car, for example, strongly indicated a possible Trump voter. 15Again, these prox- ies were sought in relation to inflection points—points at which curves would bend in new d…
41
Crucially, the predictions were trained on carefully curated data, which determined both the coefficients of the regression models and their significant components. The models were then tested on their abil- ity to predict this meticulously pruned past.
45
Using this form of verification, standard for machine learning algo- rithms and models, means that if the captured and curated past is racist and sexist, these algorithms and models will only be verified as correct if they make sexist and racist predictions, especially if they rely on prob- lematic measures such as standard I…
47
The methods used by Kosinski and colleagues and Cambridge Ana- lytica—correlation, linear and logistic regression, and factor analysis— stem from twentieth-century eugenics. The five-factor OCEAN model is the product of controversial and discredited eugenicists such as Charles Spearman, Hans Eysenck, and Raymond Cattell. They developed and used factor analysis, based initially on principal component analysis (PCA; see figure 37 by Alex Barnett in “Proxies, or Reconstructing the Unknown” after chapter 2) and correlation to “classify” raced and gendered groups according to intelligence, among other personality trai…
47
In the “five-factor” world, personality traits or factors were, and still are, considered “physiological.” According to Robert McCrae and Geert Hofstede, the “five-factor model” was “unique in asserting that traits have only biological bases”22—an assertion that provided the basis for researchers using the model to frame personality within a biometric evolutionary schema.
47
According to Christopher Wylie, to more effectively influence people, Cambridge Analytica took an “intersectional” approach to racial identity. Steve Bannon, he explained, was the only straight man he talked to about feminist intersectional theory, a metho…
49
24Cambridge Analytica perverted Crenshaw’s method and sought to augment differences and exclusions, both real and perceived, based on values and “personality traits,” within a racially homogenous space.
49
Put most bluntly: in an attempt to destroy any and all senses of com- monality, “communities” are being planned and constructed based on divisions and animosities. Instead of ushering in a post-racial, post- identitarian era, these social networks perpetuate angry microidentities through “default” variables and axioms. By using data analytics, indi- vidual differences and similarities are actively sought, shaped, and instru- mentalized in order to capture and shape social clusters. Networks are neither unstructured masses nor endless rhizomes that cannot be cut or traced. Because of their complexities, noisiness, and persistent inequali- ties, networks provoke control techniques to manage…
49
A worldview we thought was made of causes is being challenged by a p ponderance of correlations.
50
…big data posed and still poses f cinating computational problems—How do we analyze data we can read only once, if at all?—and the plethora of correlations it documents raises fundamental questions about causality. If almost anything can be shown to be real, if almost any correlation can be discovered, how do we know what is true?
51
32Thus to understand the impact of the “data deluge,” we need to move beyond celebrating or dismissing big data toward comprehending the force of its promise—or, more precisely, the ways it undermines the promise of promise.
51
Correlation’s eugenicist history matters, not because it predisposes all uses of correlation towards eugenics, but rather because when correlation works, it does so by making the present and future coincide with a highly curated past. Eugenicists reconstructed a past in order to design a future that would repeat their discriminatory abstrac- tions: in their systems, learning or nurture—differences acquired within a lifetime—were “noise.” The important point here is that predictions based on correlations seek to make true disruption impossible, which is perhaps why they are so disr…
52
As philosopher of science Ian Hacking has pointed out, the term “statistics” comes from “state,” and national statistics testify to a state’s “problems, sores and gnawing cankers.” 34Data science, in con- trast, by focusing on the governmental interests of…
52
For the twentie eugenicists, homophily was an aspiration: they wanted to create a world in which like people automatically reproduced with like. In data analyt- ics, homophily is a given, an axiom. Nightmares of global destruction and dreams of segregated “escape” have displaced narratives of impending racial doom. So how did we get here, and what is correlation a…
52
As sociologists Josh Cowls and Ralph Schroeder explain, instead of either correlation or causality alone, what is necessary are “mixed methods” that combine correlational exploratory practices with causal explana- tory research.
56
This is because, left unattended, big data methods often reinvent the wheel by “discovering” well-known latent correlations (that many gay men of a certain age like Britney Spears, to return to an example referenced earlier), or they produce an inordinate number of spurious cor- relations that defy basic concepts such as gravity or photosynthesis. Fur- ther, causality is often needed to solve problems—vaccines, for example, depend on mechanistic understandings of virus structure and beh…
56
Further, spurious correlations arrived at using big data are not accidental; indeed, drawing on mathematical theory, theoreti- cal computer scientists Cristian Calude and Giuseppe Longo have shown that, because of their size alone, all big data analyses must be riddled with such corre…
56
Correlations, again, do not simply predict certain actions; they also form them. Correlations that lump people into categories based on their being “like” one another amplify the effects of historical inequal- ities.
58
The problems with correlations are neither new nor limited to big data and weapons of math destruction, however. Based on eugenic reconstruc- tions of the past and cultivated to foreclose the future, correlation contains within it the seeds of manipulation, segregation and misrepre…
59
British eugenicists developed correlation and linear regression, key to machine learning, data analytics, and the five-factor OCEAN model, at least a century before the advent of big data.
59
C relation was never simply about discovering similarities, but also about cultivating physical similarities in order to control the future. Correlation provided the basis for eugenics’ “universal laws.”
59
As Ruth Cowan and other historians of science have shown, Galton developed regression and correlation while studying heredity in humans and plants and the identification of criminals.
59
A biometrician rather than a Mendelian, Galton believed that all traits were distributed along a normal curve within a population, rather than determined by genes.57 Exceptions, such as genius, were statistical outliers and thus located at the ends of the curve, in the fourth quartil…
60
Galton’s linear reversion thus differed significantly from the now standard linear regression. In tracking how generations deviated from the norm, his goal was to maximize “good” deviation. In contrast, lin- ear regression seeks to minimize standard deviati…
63
Linear regression is typically used to determine the best line between a scattered set of points, where “best” means the line that minimizes the distance between the data points and the projected line.
63
59In this version of correlation—a version more commonly used in statistics—correlation is used to cut down on the number of variables involved, not to uncover “hidden” or latent variables.
64
Pearson also believed both natural and artificial selection could eas- ily and continuously affect future generations: the past and future were linked linearly. In contrast, Mendelian eugenicists did not hold such a simple, progressivist view since regressive traits could reappear at any time and thus frustrate phenotype-based breedi…
64
62Mendelian eugenicists thus sought to create “pure” bloodlines cleansed of “undesirable” traits, whether dominant or recessive, whereas biometricians viewed racial or national populations as inherently mixed and intermingled; there was no
64
“pure” breed, and positive deviations needed to be preserved and dissem- inated. Eugenicists in both camps, however, held individuals responsible for the future: their behavior could either benefit or destroy the nation.63 And both camps believed that nature triumphed over nurture, making eugenics central to breeding a “better” national…
65
…t- ened to destroy the English race: through them, the “unfit” multiplied at the expense of the “fit.” 69In the nationalist view of biometric eugenics, every citizen was connected: natural and artificial selection operated at the level of the nation-state.
65
… asking: To what extent has eugenics reemerged—if it has—not simply or directly through the proliferation of genetic test- ing and manipulation, but also through biometric methods and predic- tions?72 And how have data analytics and machine learning been used to found a revised form of eugenics, in which discriminatory pasts, pr…
66
To what extent do the current descriptions of correlation as unlocking the future reflect the twentieth-century celebrations of correlation and its confidence in eugenic solutions? To what extent can understanding this mirroring help elucidate why and how the world of data analytics and machine learning, based on meth- ods arising from these descriptions, feels so small and enclosed? And how did a worldview that did not believe learning could happen— that intelligence could only be bred—become the basis for machine learn…
66
In addition to treating correlation as inherently predictive, there are many similarities between twentieth-century eugenics and twenty-first- century data analytics. Both emphasize data collection and surveillance, especially of impoverished populations; both treat the world as a labora- tory; and both promote…
66
…losely con- nected social practices are with the future vigour of the nation,” 73eugen- ics required detailed surveillance of human populations. Eugenicists thus collected data to produce charts documenting the transmission of traits (such as criminality).
67
Virginia Eubanks has linked twentieth-century eugenics to twenty- first-century data analytics and machine learning through their practices of surveillance. Eugenics she has revealed, “created the first database of the poor,” 75and contemporary programs to automate public services pro- grams have given rise to digital poorhouses, all too similar to the physical poorhouses of the nineteenth century, which imprisoned and punished the p…
67
Both eugenics and big data use surveillance in order to experiment with humans.
68
…behavior, one that can be explored, predicted, and no doubt exploited.” 82If twentieth-century eugenicists however defended their work against accusations that it experimented on humans, twenty-first- century data scientists openly embrace experimentation.
69
What is most significant, however, is that the eugenicist aspiration—the reproduction and selection of like with like—has now become axiomatic.
71
This normalization of homophily also drives the other major differ- ence between twentieth-century eugenics and twenty-first-century data analytics: the move from the nation to the neighborhood through the notion of individual preference.
72
In the world of the biometricians, members of national populations—assumed to be racially homogeneous— were inextricably intertwined: the fate of one person affected that of the others, hence the need to restrict others in order to help oneself. In the world of the “Sovereign Individual,” exit reigns supreme because, in the place of nationalism, there are “communities and allegiances . . . not territorially bounded. Identification . . . [is] precisely targeted to genu- ine affinities, shared beliefs, shared interests, and shared genes.” 99The relationship between individuals and populations still matters, but the relevant group is now the “network neighborhood”—or the homophilic cluster, groupings that are based on kinship and specialized interests rather than on notions of equal…
73
… racial under- and overtones, has not completely undermined national interventions or identity, it has meant that national interventions hap- pen through the functional equivalent of what Antonio Gramsci diag- nosed as “hegemony,” albeit formed in reverse.
73
The transgressive hypothesis
“Being red pilled” and reverse hegemony depend on the transgressive hypothesis: the notion that individual defiance and difference ground freedom.
75
Mainstreamed after World War II in reaction to Nazi eugenics and Stalin- ism, the transgressive hypothesis equated democracy with nonnormative structures and behaviors—anything but the conformity that supposedly drove totalitarianism. Its enemy was “the herd,” the basis for national eugenics and mass society; its remedies were “thinking different” and new …
76
But this constant call to be different—this mainstreaming of resistance against repressive norms—far from “automating” democracy has instead fostered populism, paranoia, polarization and the new biometric eugen- ics.
77
What is needed, in other words, is not a rebirth of the “man of action”—or his command to forget in order to control. If we are truly to move beyond our fears and obsession with a freedom that is no free- dom—if we are to halt the looming extinction of the majority of human- ity in its tracks—we need to engage the richness of what we too easily dismiss as “the past.” If the past and future are similar, it is because they are both unknown—our (re)constructions of them cannot begin to touch their richness. What potential might we find if we were simply to revisit and reimagine what has been dismissed as “traini…
80
Homophily, or, The swarming of the segregated neighborhood
By the early twenty-first century, the imaginary of the Internet had moved decisively from the otherworldly expanse of cyberspace to the domesticated landscape of well-policed, gated “neighborhoods.” This was a progression rather than a transformation: U.S. settler colonialism and enclosure underlay the visions of both neighborhoods and cyberspace.
81
Homophily is used to c ate agitated clusters of individuals whose angry similarity and overwhelm- ing attraction to their common object of hatred both repel them from one another and glue them together. Crucially, homophily stems from mid-twentieth-century analyses of white U.S. residents’ attitudes toward biracial public…
82
Katz and Lazarsfeld’s analysis insinuated that the problem in a “post-truth” world is not lack of trust, but self-doubt and subsequent group trust; in other words, it is not that people question mainstream media, but rather that, in doing so, they come to trust other, more dubi- ous sou…
84
This chapter asks: To what extent has this description become a prescription—a guide to divisively politicize majorities by undermining their ambivalent solidity if not solidarity?12 To what extent can ambiva- lence be used to diffuse polarization and provide the basis for democratic political possibi…
85
The “awakening” of “the masses” depended on polarization—on the creation of “neighborhoods”: it transformed inert nodes into clusters of “charged” elements.
85
As feminist and queer theorist Sara Ahmed has argued, to bind together “I’s” into a “threatened” “we,” hatred needs the other. 13Remarkably, this hatred is rephrased as “love,” as “homophily”: modern white supremacists, for example, claim not to hate others but to “love” their own.
85
Spaces of sameness thrive on comforting yet repelling rage.
85
Through homophily, network science and data analytics as currently configured inadvertently perpetuate the discrimination they find.
86
At the most basic level, network science captures—analyzes, articulates, imposes, instrumentalizes, and elaborates—connection.
86
Fundamentally interdisciplinary, network science brings together phys- ics, biology, economics, social psychology, sociology, and anthropology. But, in merging the quantitative social sciences with the physical and computer sciences, it bypasses the qualitative social sciences and humani- ties, fields also steeped in theories of representation and…
90
…r, one that can be explored, predicted, and no doubt exploited.” 23Network science unrav- els a vast collective unconscious, encased within the fishbowl of digital media—which is why a degree in computer science is now more relevant than one in psychology.
90
In attempting to do so, network science reduces real-world phenom- ena to a series of “nodes” and “edges,” which are, in turn, regenerated to reveal the causes of seemingly disparate behaviors, from friendships to financial crises. These “discovered” relations are vast simplifications of vast simplifications, with each stage of network theory—initial abstrac- tion or representation, followed by mathematical model…
90
its own type of abstraction.
91
This includes, for example, what constitutes an individual entity or a relationship, how to conceptualize the strength of a tie, etc.” Most simply, this stage decides what is a “node,” what is an “edge,” and how they should be mapped.
91
This second stage builds models that reproduce the abstractions produced in the first. Whatever repeats the initial mapping is true or causal: truth within these mathematical or logical systems as Arendt points out is consistency (see “The Totalitarian Hyp…
91
This two-stage process highlights the tightrope between empiricism and modeling that network science walks when it simulates abstract representations of the world and asserts that “truth” is what reproduces these abstractions. This two-stage process defines capture systems more generally.
91
Without dependence among ties, there is no emergent network structure.” 28At all levels, networks are dynamic and interdependent. What matters is understanding and creating ties.
92
Modeling these interdependencies—tying global events to individual interactions—entails coupling graph theory with game theory or other agent-based modeling protocols.
92
As the turn to game theory indicates, a market-based logic permeates network science models. Indeed, network science and capture systems are arguably neoliberal “cures” for postmodern ills.
92
In a neoliberal society, the market has become an ethics—it has spread everywhere so that all human interactions, from motherhood to educa- tion, are discussed as economic “transactions” and assessed in cost-benefit terms.
93
37Most succinctly: capture systems translate and transform all human interactions into market-based exchanges so that computerization neoliberalization.
93
…cation effects in dif- ferentiated markets.” 40As a relational form of capital, social capital grants advantage to those who are “somehow better connected” and invest in social relations; it thrives off “trust,” obligation—and location, location, location.
94
…nd material stratification”; über-capital thus “subsumes unlucky circumstance and uncaring social structure into morally evaluable behav- ior.” 44In other words, through habits—shards of others encased within the self—über-capital launders group advantage.
94
Crucially, network science and capture systems reshape the activities they model or “discover.” 46Through a metaphor of human activity as language, they impose a normative “grammar of action” as they move from analyzing captured data to building an ontological model of the captured activity.
95
48Social networks create and spawn the reality they imagine; they become self-fulfilling prophecies.
95
By implicitly validating segregation as a personal choice and by erasing institutional and economic constraints, network science inadvertently furthers racist agendas. It buttresses neoliberal and Sovereign Individual plans to destroy society by proliferating segregated neighborhoods. Net- works preempt and predict by correlating singular actions to larger col- lective habitual pat…
95
…as a “commonsense” concept that slips between cause and effect, homophily assumes and creates segre- gation. It transforms individuals into “neighbors” who naturally want to live with people “like them”; it introduces normativity within a sup- posedly nonnormative system by presuming that consensus stems from similarity; and it makes segregation the default. In valorizing “voluntary” actions, it erases historical contingencies, institutional discrimination, and economic realities. At its worst, it serves to justify the inequality it maps, by relabeling hate as “love.” When homophily, rather than rac- ism or sexism, becomes the source of inequality, injustice becomes “natu- ral” or “ecological,” and conflicting opinions, cross-racial relationships, ambivalence, and even heterosexuality …
96
The qualifications and context provided in “Friendship as Social Pro- cess” have been erased in the early twenty-first-century form of network science. Homophily is no longer a problem or a question, but rather a solution. In the move from “representation” to “model,” homophily is no longer something to be accounted for, but rather something t…
101
“naturally” accounts for and justifies the persistence of inequality within nominally equal systems. It has become an axiomatic, commonsense principle, thus limiting the scope and possibility of network science.
102
…p.” Homophily also accentuates network clustering. Although sometimes considered a structural cause different from homophily, triadic closure also presumes homophilous harmony and consensus. Triadic closure presumes that “if A spends time with both B and C, then there is an increased chance that they will end up knowing each other and potentially becoming friends” in part because “if A is friends with B and C, then it becomes a source of latent stress in these relationships if B and C are not friends with …
102
other.” 83Social networks such as Facebook amplify the effects of “triadic closure” and “social balance.” By revealing the friends of friends—and by insisting that friendships be reciprocal—Facebook makes triadic closure part of its algorithm: it is not simply predictive; it is also prescriptive.
103
84Network science posits n connection as unsustainable—a cause of stress. Conflict or indifference as ties are difficult to perceive or conceive. Homophily not only erases conflict; it also naturalizes discrimination.
103
Pen
103
Again, homophily maps hate as “love.” How do you show your “love” of the same? By running away when others show up.
105
So, what would happen if we engaged, rather than decried, social net- work performativity? How different could this pantomime called “social networks” be if we created new grammars of action by understanding how our silent—and not so silent—actions registe…
105
Academic references to homophily increased at the end of the twentieth century and the beginning of the twenty-first. Its canonization coincided with the rise of recommendation engines (“recommenders”) and collab- orative filtering. The three disciplines responsible for this marked increase were computer science, sociology, and behavioral sciences (figur…
106
These figures reveal the constant erasure of ambivalence and “weaker” affects in favor of stronger ones. Through this erasure, a more racially dichotomous situation than actually existed is portrayed. What would have happened if the numbers for friends and acquaintances had been the basis for their analysis, rather than the numbers for three closest friends? What notions of cross-racial homophily would have emerged?
117
This valorization of disruption, friendship, neighborhoods, social engineering and experimentation connects these early studies of public housing projects to the twenty-first century world of social media with its “new biometric eugenics.” It is no accident that pre–social media stud- ies of homophily focused on schools and that social media sites take the college campus as their architectural and social model. The constant disruption of habits in order to create new friendships and accentuate differences—to transform the “politically inert” or “ambivalents” into agitated partisans—echoes the Columbia researchers’ description of hous- ing projects. The move from “the mass” to the new depends on making the “ambivalent” unstable. It depends on a logic of “authenticity”—of “latent” and “manifest” features—and of “comfortable” spaces of in which “secret” racial attitudes can be revealed. Their “Patterns of Social Life” study, however, also reveals the extent to which other futures could have emerged and still can emerge, not through the suppression of indif- ference, but…
120
Proxies, or, reconstructing the unknown
Blaming proxies for race for racist discrimination presumes that racism naturally stems from visible difference—that if we just didn’t track race, racism would disappear. As many studies and the first half of this book have clearly shown, however, it would not, and, indeed, this color-blind presumption—this hopeful ignorance—is dang…
122
How do seemingly race-, gender- and difference-free models perpetuate discrimination? What do the proxies that help them perpetuate discrimination reference and do? And how can we use the results of these models “against the grain,” as evidence of discriminatory p…
122
What would happen if we treated these and other models as we do global climate change models? Climate models predict the most probable future for the world, given past and current actions, not so that we will fatalistically accept the future they predict, but rather so that we will do whatever is needed to prevent that future from occurring. 4Being accurate in the narrowest sense of this word—encouraging us to keep producing hydrocarbons so that we verify their predictions—is not the point. When a global climate change model we have good reason to trust predicts a rise of two degrees Celsius in the global temperature, we seek to fix the world, not the model that predicts them—unless, of course, we are global climate change deniers.
122
This analogy to global climate change models also reveals the limits of models and explanations.
122
The example of global climate change models also complicates cri- tiques of proxies—it points to their necessity and to the political struggles they inevitably evoke. I therefore place social networking algorithms next to global climate change models to make us pause: to shake loose our nor- mal assumptions and conclusions and, in particular, to make us recon- sider blanket critiques …
123
Proxies both reduce and introduce uncertainty. By representing the unknown or absent, they evoke the specter of the unknowable.
125
In effect, principal component analysis breaks data down into a set of vectors to reveal significant patterns: the first principal component will explain the greatest variation since most of the data lie along that
127
More simply, principal component and eigenvector analysis recenter data around a new set of axes, which makes mathemati- cal calculations much easier.
128
Proxies are not inherently “innocent” but neither are they inherently “guilty.” They are central to both understanding global climate change and to creating “weapons of math destruction.” When used to seek the unknown or absent, they introduce uncertainty, even as they serve to reduce it. Proxies are necessary and inadequate: indeed, they point to inadequacies in direct knowledge more generally.
136
A proxy embodies what Jacques Derrida called a “pharmakon,” a supplement or intermediary: “a philter, which acts as both remedy and poison.” 29Proxies absolve one of respon- sibility—a payment in lieu of hospitality—by creating new dependencies and relatio…
137
Proxies touch the unknown: they extend the knowable, by capturing or “syncing up with” what is not there. Proxies spark contro- versy and raise questions about the relations they “uncover.”
137
Algorithmic authenticity
As the question “Does Donald Trump say what he believes most of the time, or does he say what he thinks people want to hear?” implies, authen- ticity in this political sense entails a person’s vocal disregard for conven- tion: a “subversiveness” that indicates that a person has neither restraints nor filters. As this chapter reveals, the “subversiveness of authenticity” drives predictability by urging users to make their outer and inner selves coincide. The constant call to reveal inner secrets or to transgress against the mainstream dispels ambiguity, heightens affect, and valorizes behav- ioral transparency. “Recommender” systems and social…
139
as well as more mainstream media forms such as reality TV, have opera- tionalized authenticity—the imperative to “be true to oneself”—in order to provoke predictable responses to their prompts. Authenticity however entails drama and participation. We are characters and actors—not mari- onettes—in the drama we so inadequately call “b…
140
Comparing The Apprentice and Trump’s campaign reveals that, rather than running off script, Trump was unwaveringly and single-mindedly on script—a script that had begun as early as 2004. Through this unending performance, Trump managed to make all other candidates appear to be two-faced hypocrites because they did not host a reality TV show.
143
So how did “Reality TV” and someone so staged—so unabashedly a brand—come to define authenticity?
144
And how did authenticity become redefined as the expression of “subversive” opinions—opinions that were or would quickly become dominant on the social media and networks?
144
Trying to determine whether Trump—or, indeed, any figure or brand— is really authentic will not answer these questions. Such attempts are at best decoys that distract us from seeing the historical and well-known paradoxes that structure the concept of authenti…
144
Authenticity is an ethos used to evaluate social performance—to “authenticate” and mold “good” users—especially when they are “bad.” It prescribes a certain transpar- ency of self that makes someone’s data reliable. It is the flip side of con- formity or sincerity: if we conform by making our inner selves coincide with our outward appearance, we become authentic by making our outer selves reflect our inner torment. Either way, we become “tr…
144
Further, authenticity has become so central to our times because it has become algorithmic (if it was ever not algorithmic): a set of rules to be followed or executed. At its very simplest, authenticity evokes the (dra- matic) command: “To thine own self be true.” To call authenticity algo- rithmic, however, does more than highlight the methodological nature of authenticity. The term “algorithmic authenticity” reveals the ways in which users are validated and authenticated by network alg…
144
Fundamentally about recognition, algorithmic authenticity buttresses human and machinic pattern recognition. It ties together supposedly separate—or even competing—agents and platforms. It underlies person- alized recommendation engines, social media, and network clust…
145
According to Trilling, the rise of authenticity coincided with the fall of sincerity during the nineteenth and twentieth centuries. By the countercultural 1970s, authenticity clearly dominated over—and was used to evaluate—sincerity: a person’s ability to appear sincere was judged by its authenticity…
146
Self-branding as authentic represents another twist in the history of authenticity: it has become openly relational. If, as Trilling argued, sincer- ity and authenticity once seemed to differ in their approach to the adage “To thine own self be true”—to be “sincere,” you are true to yourself in order to be true to others, whereas to be “authentic,” you are simply true to yourself—branding authenticity amplifies authenticity as a form of self or subject possession (through your authenticity, I become authen- tic; and through my identification, you become…
148
Thus, through this carefully crafted and scripted visibility, authentic- ity has come to defy definition.
148
That a politics of authenticity favors a reality TV actor is thus to be expected, for reality TV is strictly formulaic, and it blurs the difference between actor and character. As well, the ties between reality TV and neo- liberal governmentality run deep. Thus, in many ways, Trump is a train- ing program for t…
149
…the emergence of reality TV as a popular television format coincided with the dismantling of the last vestiges of the welfare state within the United States. Reality TV shows from Extreme Makeover: Home Edition to What Not to Wear to The Biggest Loser have emphasized individual empowerment and responsibility, as well as corporate charity and the reinvention of government…
149
Reality TV is a program, in all senses of that word.
150
If the “proof” of the Internet’s libertarian d cratic nature lies in its diversity, everything must be displayed—offensive content must be generated—in order for the Internet to be deemed suc- cessful. In this worldview democracy = offense. Further, this constant display of everything—and, in particular, of minor “deviations”—is used to authenticate and cluster users. The demand to be authentic—to trans- gress convention and the boundary between public and private in as scripted a manner as possible—makes data “the oil of…
152
There is nothing particularly human or particularly machinic about either algorithms or programs. So how exactly does human execution differ from machinic execution? Surely, reality TV with its formulaic yet unexpected outcomes—the conflict, horror, disgust, fascination, laughs, and surprise it generates—differs from machinic commands and performance? Yes, it does in many ways, but both produce unexpected results or they would not be necessary, for we would know those results/
152
outcomes in advance.
153
The difference between human and machinic execution lies in how humans and machines are trained and “authenticated”: how the wildly continuous nature of signals and persons alike is made discrete and
153
molded; how patterns are recognized and fostered. Recommender sys- tems and the types of “personalization,” enabled by network algorithms and supposedly driven by a desire for mutual human and machine “learn- ing,” make this point clear.
158
…most recommender systems are hybrid. They are also divided into memory- or model-based systems. Although both of these rely on past interactions, memory-based systems hold all (or most) data “in memory” and use them directly to generate recommendations (which is computa- tionally intensive, especially at run time), whereas model-based systems preprocess the data and then use the “learned” model to make rec…
159
That recommender systems are based on homophily raises several important questions: What are the ramifications of similarity? How do we measure it? And how are items and users determined to be similar?
159
Regardless of the methodology used, “collaborative” recommender systems generally tie together the past, present, and future through “link prediction,” which, as computer scientists Jun Zhu and Bei Chen explain, is “one of the most fundamental problems in network analysis.”
160
As computer s entists Dietmar Jannach and colleagues explain, sites deploy well-known priming tactics to make certain purchases or items more attractive, such as adding “irrelevant (inferior) items in an item set [to] significantly influ- ence the selectio…
160
These “collaborative” systems also carve networks into affectively intense “neighborhoods.” They create these neighborhoods by clustering users who deviate from the norm, the mean, the common denominator with fellow members of their neighborhoods.
160
62The point here is to find and amplify triggers that ensure predictable—linear— user reactions and that can be used to delineate the boundaries between polarized neighborhoods.
161
The move toward latent factors, which is historically linked to Lazars- feld and Merton’s 1954 work on value homophily as underlying status homophily (see the following “Correlating Ideology or What Lies at the Surface” section), raised the possibility of uncoverable “causes,” which could then be exploited. Latent factors, computer scientists Animashree Adnandkumer and colleagues explain, are “central to predicting causal relationships and interpreting the hidden effects of unobservable con…
162
The problem with these and any learning-based system, of course, is the creation of “more of the same.” 69All these systems—whether they use nearest neighbor methods, matrix decomposition, or neural networks— restrict the future to the past. They are successful because they “recom- mend” things that are immediately recogni…
163
Contrary to much hype about big data, measures of similarity are neither theory- nor bias-free. The continuing use of the Pearson correlation coeffi- cient points to the enduring legacy of eugenics, just as homophily indicates the continuing impact of segregation and social engine…
165
Biometric e ics, as discussed in chapter 1, assumed that all correlations were due to “nature” and that all generations “regressed” to an ancestral mean, unless positive deviations were carefully propagated through sexual selection. These correlations thus were key to developing programs to breed better populations.
165
For these recommender systems to work, though, users have to become predictable subjects: they must be authenticated and determined to be operating authentically. Recommender system programs presume that users’ captured actions—rather than their deliberate speech—represent their true selves; hence the claims made by data scientists to have “cap- tured” users’ true and nonconscio…
165
In general, these systems p sume that when people are at their most emotional—when they are the most agitated or distracted—they are most “truthful,” and they act most predictably. The “good” user is an authenticated one, who acts indepen- dently and whose collaborative actions are “accid…
166
In “collaborative” recommender systems, deliberate collaborative actions are framed as “inauthentic” attempts to break the systems. Iden- tity politics becomes “inauthentic” and “noncollaborative” and solidarity becomes “disingenuous.” These systems, in other words, presume neo- liberalism: that the world is filled with competing individual agents and that to act collectively—to make conscious collaborative connections with others—is to “game” the …
167
R ers and reality TV reduce authenticity to transparency or “sincerity”: to be “genuine” is to be consistent—without intention or design.
168
If we think through our roles as performers and characters in the drama called “big data,” we do not have to accept the current terms of our deployment. Indeed, by acknowledging and engaging the wonderful creepiness of social networks, we can replace this big data drama with another, in which we take on the myriad and constant actions necessary
170
to maintain those networks. We can move from tragedy to comedy or fantasy. The goal before us is to move the big data drama away from preemption and predictable yet rampant consumption toward political contestation and sustainable habitation.
171
Correlating ideology, or, What lies at the surface
To repeat: How might correlations become the basis for a new collective politics? How might they open—rather than close—the future? To answer these questions, we need to understand how authenticity and correla- tion lie at the surface. Authenticity—as a twenty-first-century branding technology—demands “transgressive” transparency: boundless selves with no secrets, with no faces to save. Correlation and authenticity lie on the flip sides of the sam…
173
For authenticity to be transgressive, the real world must be latent, beyond view, lying in wait—beneath the surface. Correlations reveal proxies; they transform the unknown into the knowable. They make manipulatable “latent factors” manifest, from personality traits to microgenres.
174
Althusser stresses that these hailings “hardly ever miss their man: verbal call or whistle, the one hailed always recognizes that it is really him who is being hailed.” 25Ideology is a communicational event, in which response = recognition.
181
26Without operating systems, there
181
would be no access to hardware—indeed, without them, there would be no actions, no practices, no users. Each operating system, through its brand, calls to its “users” and offers them a name or image with which to identify.
182
And why is it that software seems to p fectly mimic every definition of ideology we have?
183
…reconstructed not by a process of piecemeal decoding, but by the identification of the generative sets of ideological categories and its replacement by a different set.” 31Ideology is not simply illusion or fiction, but rather “real” correlations that lie.
183
Recognizing recognition
As this chapter reveals, the links between facial recognition technology and eugenics are not only thematic or aspirational, but also methodological. They are rooted in eugenic meth- ods, such as linear discriminant analysis, developed in the early twentieth century to discriminate between classes and races of peo…
186
…this chapter analyzes FRT in the context of (1) cybernetic attempts to analyze humans, machines, and animals as enduring and generalizable patterns; (2) the “recognition versus redistri- bution” debates of the mid-1990s; and (3) the rise of the “new politics of recognition” by the reactionary right in the early twenty-first century as a way to prevent economic and political redistribution. Through these developments, recognition has become a metaphor for discrimination: like homophily, it launders hate into “love.” However, as the fears about “AI masters” that haunt machine learning dreams indicate, the issue of recognition is no trivial…
186
To recognize someone is to accept that person’s authority, validity, or legitimacy.
186
According to Wang and Kosinski, machine learning outperformed humans (aka U.S. Mechanical Turk workers) in correctly reading a person’s sexual orientation. Colloquially called “machine gaydar” by journalists, their model scanned white U.S. faces to in order to predict their sexual orien- ta…
186
Significantly, the researchers did not report accuracy percentages for determining the sexual orientation of the owners of faces displayed in the images when comparison images were not used.
188
Since the operations of the deep neural network model were opaque to the researchers, they engaged in a series of “hacks” to decipher what mat- tered. To discover what the model was actually reading—to make sense of the latent dimensions they had constructed—they randomly chose 100 face images of men and 100 face images of women for further analys…
188
Jawline, nose, forehead, facial hair, skin color, eye makeup, and gender-typical grooming—these were the proxies that mattered.
190
The relentless framing of recognition as a choice between a gay or lesbian and a straight face image not only erases gender and sexual ambiguity and transgender and transsexual subjects, it also raises doubts regarding the generalizability and relevance of their study’s results.
192
The researchers relied on a “biological” theory of sexual orientation to justify their race-based exclusion. If sexual orientation is caused in utero, what holds for one race, they asserted, should hold for all races.
192
The links between eugenics and recent studies on facial recognition technol- ogy are not only topical or aspirational, but also methodological. Princi- pal component analysis (PCA), developed by eugenic biometricians (see chapter 1 and “Proxies” after chapter 2), drove one of the most significant late twentieth-century advances in FRT: the eigenface method, which moved facial recognition technology away from humanly determined toward nonhumanly—algorithmically—determi…
194
Like Wang and Kosinski, Galton argued that, through technology, he could recognize legible characteristics not normally discerned by unaided human observers. 17Galton called his images “composites” because he asserted that they were visual equivalents to Adolphe Quetelet’s statistical averages, even though his sample size did not approach the size used to produce statistical tables.
194
Facial recognition technology and modern computationally based biometric techniques have merged Bertillon’s and Galton’s projects. By “authenticating” both individual and type, they seek to produce “authenticity machines.” Although most pattern recognition…
196
The desire to “read” images—to archive them and deploy them in order to identify both the particular and the generic—drives the transformation of discriminant mathematical func- tions into algorithms of “recognition.” But how did identification and discrimination become recognition? And why does this mat…
198
As the term “linear discriminant analysis” (LDA) implies, Fisher devel- oped LDA functions to discriminate between races and species. By dis- cerning and using the “measurements by which the populations are best discriminated,” these functions built mathematical fences between popu- lations whose boundaries app…
198
Before there was pattern recognition, there was pattern discrimination. “Discrimination”—the ability to divide, separate, and distinguish—paved
205
the way historically and theoretically for “recognition.” The “gaydar” example that started this chapter nicely reveals the enduring ties between the two for “recognizing” gay or lesbian faces meant first distinguishing between two types: “gay” or “lesbian…
206
Categorization underlay domination. Control systems were also called “servomechanisms”: devices that controlled the message or pattern through enslaving, or enslaved through controlling. Thus it is no surprise that pattern recognition became a founding task of artificial intelligence.
211
Early on, artificial intelligence focused on producing machines that could “recognize” patterns, that is, record similarities across different contexts. In 1960, pattern recognition was framed as one of the most important tests for intelligence that machines had not yet passed.
211
According to Sayre, humans could identify x as a member of a class, without being able to specify the characteristics by which they did so. But because machine recognition could not follow the path of human intuition, it had instead to determine invariant characteristics, “eas- ily expressible in computer language,” that would distinguish “a given class of individuals from all other classes, and the possession of which qualifies an individual for membership in that class.” These invariant fea- tures would serve as proxies for human intuition and would “enable the computer to select approximately the same inscriptions which a human typically would recognize as members of that …
212
M ern pattern recognition systems draw together various disciplines and approaches to discrimination. They translate between the statistical, syntactical, and cognitive. They equate inference with generalization; learning with estimation; and classification with discriminant analy- sis. Classification systems require the prior construction or discovery of “invariant” features, on the basis of which they assign and reduce obj…
212
context, the most discriminating).” 61The most discriminating features are valued, regardless of their physical, biological, or conceptual impor- tance—recognition = discrimination++—for they seem “invariant.”
213
Mimicking queer families formed in response to homophobia, the reactionary right creates new militant “tribes” from the ashes of civil society norms. Crucially, these tribes not only define themselves against mainstream culture; they also embrace and amplify a perceived stigma: they identify as militant victims, hence their attraction to queer and black liberation methods.
217
…dis-identification: the subject “neither opts to assimilate within such a structure nor strictly opposes it; rather, dis-identification is a strategy that works on and against dominant ideology. . . . This ‘working on and against’ is a strategy that tries to transform a cultural logic from within.”87 Dis-identification fosters an enabling misreading: subjects read them- selves and their own life narratives in moments, objects, or other subjects that are not initially culturally coded to “connect.” This is arguably what alternative influence networks do well—they disidentify as vi…
217
This would seem the classic case of “mis-recognition” diagnosed by Taylor: through physical characteristics, incels “recognize” themselves as forever damaged. 97The difference, though, is that this inferiority complex—“biology as destiny”—is seemingly self-imposed: eugenics is wielded not to denigrate an other, but the self.
221
The point, however, is never to be a “normie,” but rather to be rec- ognized as exceptional: as an incel or a Chad.
224
They arguably hate feminists as they hate themsel they are also envious of feminists’ perceived “freedom.” They are a classic case of liberation envy. In this sense, they are not masochists, but sadists.
225
Thus the incels’ seemingly irrational yet “logical” attachment to a perceived stigma—their insistence that they are not “normies”—is key to understanding the power of subcultural references and identifications within the new politics of recognition.
225
103The very term “punk,” Hebdige says, with its “derisory c notations of ‘mean and petty villiany,’ ‘rotten,’ ‘worthless’” exemplifies a process of “ironic self-abasement.” 104Militant subcultures do not cover over stigma—they flaunt it.
225
Framing it as the absent present within punk culture, Hebdige called reggae a “black hole around which punk composes itself.” 108Punk and reggae were linked metaphori- cally: punk gained meaning by putting reggae under the bar; through punk, reggae resonated with people everywhere (for more on this, see “Correlating Ideology”). Punk and other white working-class British sub- cultures drew inspiration from reggae’s historic icons of revolt—“rastas” and “rude boys,” “gunfighters” and “tricksters”—as well as from black antislavery and anticolonial strug…
226
Like punk, the new politics of recognition clearly draws from black liberation and civil rights movements, from the plot and structure of The Matrix, and also from declarations of white ethno-nationalism.
226
Alienation, like with the Rastafarians Hebdige describes, becomes utopian exile, and Black Power, denuded of its call for redistribution and reparations, becomes a way to justify “tribal” dreams of exit.
228
To “recognize” is to identify “something that has been known before.” It is to perceive someone or something as the same as someone or something previously encountered or known, or to “identify from knowledge of appearance or character, especially by means of some distinctive feature.” For a machine or computer, it is “to identify auto- matically and respond correctly (to a specific feature, object, or event).” Recognition thus always implies a historical relation and response—and power. To recognize is to reinvestigate, to become reacquainted with and to accept the “authority, validity, or legitimacy” of another’s cl…
228
I tions are “co-relations” that reveal both similarities and differences.
229
The space between us
Net-munity calls on us to engage neighbors and relationships in all their rich ambivalence. Perversely, the logic of social networks spreads the name “neighbor” everywhere, in order to impoverish it conceptu- ally. Neighbors are not innocuous—the term “neighbor” literally recalls “boors.” They are nosy and noisy. They provoke hostility, resentment, and ambivalence. They intrude, even—and especially—when they are inert. They offer, however, a way to reside in difference and to engage relations that go beyond homophily: not just heterophily, but also ambivalence and neu…
236
16To put it in poet and p losopher Édouard Glissant’s terms, neighbors are opaque and obscure; their nontransparency, however, does not hinder but rather enables rela- tion.
236
Because public space does not rightfully belong to anyone, because this space cannot be reduced to the dominant opinion that may emerge from it, it guarantees democracy: “power becomes and remains democratic when it proves to belong to no one.”18
236
Pen
236
…is what the revolutionary tradition represents.” 19Freedom—the space between us—is not the lack of relation, but the very possibility of it. Free- dom for Nancy can be both good and evil, for freedom entails a decision: furious devastation or finite space.
237
Spacing as freedom, however, raises the question of space itself and how it is emptied: once again, colonial dreams of the new. But, as Orlando Patterson, Hegel, and those suffering from liberation envy have implied, voluntarily or not, freedom as an experience—a testing of some- thing real—emerges from the oppressed. The space between us is neither “white” nor “blank,” but teeming with those whom the archive seeks to for…
237
Coda : living in/difference
Of the many observations presented in Discriminating Data, I want to underscore eight in particular:
242
1. Freedom is only meaningful if it is freedom for all.
242
2. Reducing truth to consistency forecloses not just the present and the past but also the future.
243
3. Majorities are now formed by disintegrating dominant groups into angry minorities, by divining and amplifying perceived stigmas, and by then consolidating them together around a common “enemy.”
243
4. “Authenticity” as it is now understood renders humans as predict- able as trees—but both are far more complicated than linear models presume.
243
5. Big data is the bastard child of psychoanalysis and eugenics.
243
6. Correlation is a “co-relation.”
244
7. The past is as complex as the future.
244
Our current archives can never serve as “ground truth,” for they are limited and biased. In the world of machine learning, ground truth=deep fake.
244
8. Comfort and care are not comfortable.
244
Discriminating Data’s main goal has been to get us to step 5 of my five-step program: “Draw from struggles for and practices of desegregation and equality to displace the unjust eugenic and segregationist defaults embed- ded within current network structures and to devise different algorithms and modes of veri…
244
So, what would it mean to start from these teeming spaces of diverse interactions? To make civil rights movements the ground truth of social network analysis? To fol- low Ruha Benjamin’s call to pursue and acknowledge abolitionist toolkits?
246
To live in difference, we need to start from conflict—rather than run away from it. Conflict, not hate or love, drives democratic strug- gles. Acknowledging conflict, however, does not mean amplifying it, but rather seeking modes to repair and reciprocate.
247
To engage these rich relations, we must start from the insight that the network is everything outside the graph.
248
17These moves to “ ationalize” and “visualize” outline what can and cannot be translated into networks as they currently exist and thus the continuing necessity for interdisciplinary methods and cooperation.
248
Big Data—in its most popular current form as a glorified form of data and network analytics, used by corporations such as Netflix, Target, and FICO—mines our data not simply to identify who we are (this, given our cookies and our tendency to customize our machines is very easy), but to identify us in relation to others “like us.” Our scripts, our lines, are constantly impacted by the actions and words
250
of others, whom we are constantly correlated with and unconsciously collaborate with.
251
Again, it is both disturbing and revealing that methods developed for eugenics—a system that did not believe in learning—now form the basis for machine learning.
253
…have generated the data and optimiz- ing the search for that function as much as possible.” How much do we lose if this becomes our definition of learning—especially given the limita- tions of artificial intelligence that Meredith Broussard has documented?
253
25 How might we learn from machine learning by unlearning its predictions?
253
Machine learning and predictive models as they currently exist can also resist reduction, but only if we treat the gaps between their results and our realties as spaces for political action, not errors to be fixed. We need to treat these models as we do global climate change models. GCC models offer us the most probable future, given past actions, not so that we accept that future, but so we work to change it.
254