Technologies of Speculation: The Limits of Knowledge in a Data-Driven Society — Sun-Ha Hong (NYU Press, 2020)

Starting from the question, “What counts as knowledge in our data-driven society?”Sun-Ha Hong argues that technologies of datafication are not only transforming the conditions for producing, accessing, wielding, and validating information—they are also articulating a new set of normative demands: that we, as liberal subjects, come to know our social/political world and ourselves within that world through these technologies, regardless of the technologies’ imperfections and contradictions.

Hong maintains that we are called on to live, reason, and act through ‘data-driven insights.” Confronted with technological systems of power and control so vast and complex that “transparency” becomes burdensome rather than empowering, the public becomes overwhelmed by surveillance dragnets and driven to speculation and paranoia. At the same time, provided with a surfeit of recursive data about our desires, habits, and even our bodies, we are encouraged to privilege machinic forms of intelligence and analysis over human experience and affect. In so doing, we’re rewarded for tailoring our behaviors to be compatible with the machines surrounding us—and thus compatible with the institutions behind the machines. For Hong, “datafication is turning bodies into facts: shaping human life, desire, and affect into calculable and predictable forms, and doing so, changing what counts as the truth about those bodies in the first place” (4).

Operating at the nexus of sociology of knowledge, history of science, and critical data studies, Technologies of Speculation grounds its examination into the epistemological consequences of datafication by focusing on two sites: the 2014 Snowden affair (revelations of mass-scale government surveillance) and automated self-tracking (e.g. Fitbit). Though these research sites seem distant at first glance, Hong uncovers remarkable commonalities between state surveillance on the one hand and self-surveillance on the other. In both, the process of turning bodies into facts results not in crystalline objectivity, but in troubling gaps and asymmetries: “If data expands the vistas of human action and judgement, it also obscures them, leaving human subjects to work ever harder to remain legible and legitimate to the machines whose judgment they cannot understand” (7).

So how does this often-philosophical investigation of data-driven, algorithmic intelligence and its discontents relate back to our realm of text analysis? We might start by considering how powerful regimes like the ‘intelligence community’ confront uncertainty when saturated in data. In Chapter 5, “Bodies into Facts,” Hong explores the fragile and improvisational techniques that are used to fabricate so-called “insights” out of massive reams of data and metadata. Attending to public discourse on post-9/11 counter-terrorism and surveillance, he argues that beneath the guise of objectivity, datafication is implemented through “inconsistent and locally specific strategies governing how uncertainties are selected and legitimated as predictive insights” (115). Hong specifies three techniques by which fabrications achieve knowledge status:

• Subjunctivity: “As-if” reasoning, where a hypothetical or unproven situation is operationalized for decision-making (e.g., Acting to prevent the terrorist threat that could have happened; Acting as if you are being surveilled).
• Interpassivity: Something or someone else knows or acts for the subject, in their stead. A kind of subjective outsourcing whereby one is not responsible for a statement/belief (e.g., a rumor heard elsewhere) yet that belief/statement—neither disavowed nor claimed as one’s own—nonetheless provides the basis for actions or opinions. But the other that knows and acts for the subject is not just other people—it van also be a machine.
• Zero-Degree Risk: The dilemma of calculating risk out of uncountable streams of data with the goal of total de-risking: preventing a single outcome (e.g. a terrorist attack). Statistics, probabilities, and equations are invoked to mitigate fundamental, radical uncertainty and legitimate costly regimes of surveillance and control (NYPD counter-terrorism budget and the attacks that ‘might have been’).

Though Hong focuses explicitly on terrorist threats and state surveillance, we can consider how these techniques of fabrication may be operative in computational text analysis. For instance, “subjunctivity” undergirds many classificatory schemes. Cameron Blevins and Lincoln Mullen’s package  for inferring gender based on historical first name data works as if gender is a binary, a limitation they explore in depth within their article introducing the method (Blevins and Mullen). More fundamentally, tokenization processes generally require us to treat texts as if they lack punctuation or line-breaks. We must make locally specific choices as to whether words that are capitalized should be treated differently, which will then reshape the predictive insights we can make about frequency, association, topics, and sentiments. Similarly, the process of training a model might be an instance of “interpassivity” when that model is working on a training set that has not been annotated by our own hand. In turn, we encounter issues analogous to risk calculation when weighing questions of model fit and overfit.

But given the scant references to “text” qua text in Technologies of Speculation, it seemed to me that the monograph provided few, if any specific, new avenues of research or new methodologies we might employ in text analysis. Instead, Sun-Ha Hong asks us to step back and consider the moral dimensions of the sort of knowledge production in which we are involved. He writes in the acknowledgements that “this book is about values: the subservience of human values to the rationality of technology and capital and, specifically, the transmutation of knowledge from a human virtue to raw material for predictive control” (201). From this standpoint, methods that may have once appeared value-neutral become morally charged. The study of authorship and style through quantification—sentence length, punctuation, vocabulary size—submits a humanist endeavor to machinic rationality. It isolates “style” from meaning and social/political context, reducing a lifetime of thought to numeric signature, eschewing genealogies of interpretation and contestation among human readers while at times obscuring its own algorithmic innerworkings. Consider another case: binary classification using TensorFlow neural networks. Let’s say we wanted to take a large corpus of prose paragraphs and sort it into “prose poems” and “not prose poems.” We have training data: a whole bunch of prose poetry ranging from Baudelaire to Rosmarie Waldrop, intermixed with paragraphs from magazine pieces, novels, and other texts. But after we train the model, we cannot ask it why or how it chose to call one paragraph a poem and another not. Its reasoning is obscured, and in Hong’s framework, the knowledge of what constitutes poetry has been transmuted into “raw material for predictive control,” where before, for better or worse, it was thought “a virtue.” Perhaps this is all beside the point, but at a time when humanism can feel under threat, where we are hailed by machine intelligence to shape our behaviors according to the whims of the algorithm, the stakes of computational text analysis can feel heightened, even dangerous, after reading Hong’s book.

Works Cited

I have read the first three chapters of Artificial Unintelligence. Broussard discusses the issues in applying computer technology to every aspect of our lives and the limitations of artificial intelligence. She gives the name “Technochauvinism” to the idea that computers could always solve all human problems. The example she writes about AlphaGo is pretty straightforward, telling us computers work very well or “intelligently” in a highly structured system, just like the “Smart Games Format.” However, AI or computers can lead to algorithmic injustice and discrimination due to the data or orders the creators feed them.

Chun discusses the concepts of “correlation,” “homophily,” and “network science” in her article “Queerying Homophily” in Pattern Discrimination. I also found her explanation in this interview, Discriminating Data: Wendy Chun in Conversation with Lisa Nakamura amazing; she explained how correlation is treated by eugenics and big data as a future predictor, which actually turns out to close the future. I understand what she means by “closing the future” is closing alternatives for the future that are not predictable by the current and past data. I understand homophily is something amplified by people sharing their previous knowledge, experience, behavior, preferences, etc. Echo chambers and bubbles are generated by this homophily data collection and analysis method, further reinforcing the segregation in our society.

I am trying to make a connection here between the above two readings and the Topic Modeling articles for this week. My question for topic modeling will be how to avoid oversimplifying the context, figurative language, and nuance in discovering the topics. If an alternative reading is possible, how can this kind of representation be included in deciding the topics? For example, while the Latent Dirichlet Allocation (LDA) method could give you the percentage of possibility in similarities for topics, is it possible to address something like the impossible topics or the farthest options, or any other kinds of correlations?

The readings for this week remind me of a conversation my friend and I had a long time ago. We think math is a very romantic field because it starts with an agreement between two people that 1 + 1 = 2. If we don’t have a mutual agreement like this, the world will change accordingly. For example, I read that someone was trying to train ChatGPT that 1 + 1 = 3, and the machine will eventually accept it after a bit of struggle. To summarize my post, I believe we need to revisit concepts like “homophily” and “correlation” to keep space for agreement and mutual understanding, going beyond the aim of finding superficial similarities.

The ones who are invited or planning to be in the buffet must have the idea of what is served on the menu. The background knowledge that is what I am referring to. The one perhaps is more important than the computational process or even the decision making process of how many latent topics are waiting to be discovered within a seemingly large pool of documents aka corpus. The magic is fascinating but overwhelming if the magician can not communicate with the audience. In the case of topic modeling, the magic is the machine with computational ability that does not have the ability to express on its own(not sentient!!!). There is a role for a magician, an expert to make the magic audience enchanting, putting context to the result.

Let me provide an example. A while back, I conducted an experiment of topic modeling on the DigitalNZ archive of historical newspapers. The result is the following interactive illustration of topics. I decided to uncover 20 topics that are more prevalent during the New Zealand Wars in the 1800s.

The interactive visualization is available in the following URL

https://zicoabhidey.github.io/pyldavis-topic-modeling-visualization#topic=0&lambda=1&term=

I played the role of the magician to demystify the result and present it to a broader audience who is not a historian by no means. I used intuition to and lent the superpower of Google to support my intuition to derive the bowl that represents the topic. Below is the result I came up with. I could not produce all the 20 topics I was hoping to find.

After working long hours on this project, although I am super delighted that I have produced an interactive visualization of the mallet model which is hard to produce, there is always a feeling of disappointment that I did not have the knowledge of the historian. A historian with special knowledge of New Zealand’s history might have judged better.

Attending buffet without the knowledge of the menu is like sailing a boat without the compass but isn’t it what the distant reading is? A calculated leap of faith.

Can AI accurately determine gender when we conduct different kinds of studies? Can AI help in improving gender and racial equality? Allen Jiang’s article is very clear and compelling, showing us an approach to detecting gender with a machine-learning model with Twitter data. I appreciate Jiang’s methodology and explanation but still doubt the initial questions “Can you guess the gender better than a machine” and “What are business cases for classifying gender.” From my perspective, for example, pregnancy is not a gender-specific topic, and optimizing advertising costs could be achieved by non-binary customers’ gender detection. These questions reflect the definitions of gender I constructed derived from readings in week 2, Language and Gender by Penelope Eckert and Sally McConnell-Ginet, and We Should All Be Feminists by Adichie, to name a few. Sex is undoubtedly a different concept compared with gender. Applying a spectrum of gender in a scalar structure would challenge Jiang’s model, particularly in terms of data organization.

Specifically, Jiang primarily selects Twitter accounts belonging to famous people. In the data collecting and cleaning process, how could Jiang avoid the “celebrity effect?” I could think of methods like improving data diversity. And I am also very confused why Jiang included the data of followers. Is it possible that followers could be a feature impacting gender detection or a confounding variable? I previously raised the question about the significance of repetition in experiments to reduce error effects, which assumes a single correct result as the goal. This article reminds me of cross-validation’s importance in training models’ performances in different scenarios. The key question here I propose is to reduce the dependence on a single model designed based on one dataset with specific features.

The Gendered Language in Teacher Reviews focuses on a more diverse group sharing the same occupation, “teacher.” Upon revisiting the process, I discovered that he writes, “Gender was auto-assigned using Lincoln Mullen’s gender package. There are plenty of mistakes–probably one in sixty people are tagged with the wrong gender because they’re a man named ‘Ashley,’ or something.” The data source of Lincoln Mullen’s package is “names and dates of birth, using either the Social Security Administration’s data set of first names by year of birth or Census Bureau data from 1789 to 1940.” (https://lincolnmullen.com/blog/gender-package-now-on-cran/) Unfortunately, Ben Schmidt is no longer maintaining the teach reviews site, but updating this data visualization remains crucial, as the association between names and gender conventions is constantly changing. Using data from the 1940s to train a model for detecting gender in contemporary society may not yield ideal results.

I continued to read two articles to help me think about the question bias in AI. One is an interview Dr. Alex Hanna received about her work at Distributed AI Research Institute on AI technology and AI bias and constraints. I recommend this interview, especially for her answers to the question on independent AI research and purposes not funded by tech companies.(https://www.sir.advancedleadership.harvard.edu/articles/understanding-gender-and-racial-bias-in-ai)

There is also another super interesting but quite dangerous reading named Natural Selection Favors AIs over Humans (https://arxiv.org/abs/2303.16200)

The author Dan Hendrycks is the director of the Center for AI Safety. He discussed the natural selection and the Darwinian logic applied to artificial agents. In such large-scale computational studies and models, if biased natural selection is also incorporated, would the consequences be unbearable for us as humans? If there is natural selection within AI, where does the human position be?