Sipping From the Firehose of Digital Data
Scott Weingart sees algorithms as prostheses
My guest this month is Scott B. Weingart, the director of the Navari Family Center for Digital Scholarship at the University of Notre Dame. Our conversation is prompted by Kyle Chayka’s essay “The Uncanniness of Algorithmic Style.” Kyle is also the author of The Longing for Less: Living with Minimalism.
Leah: Scott, I’m delighted to have you as my guest for this discussion. Kyle’s piece is about how machines view the world—and I’m over-anthropomorphizing here. In fact, AI oddities are often about how we see the world, and what it looks like to take those heuristics too far.
The examples Kyle highlights are funny, because we can clearly see the gap that leads a machine to assume, from the height of the Washington Monument, that it must be an office building. But the biases of a sentencing algorithm are a lot more serious, and the deviations from justice are harder to spot and correct than the mismappings.
As someone who works in digital humanities, how do you think about your relationship with algorithms? How do you make sure you aren’t amplifying blind spots? (And, if you wouldn’t mind, what are the digital humanities?)
Scott: Oh no, you asked the “what are digital humanities” question! You know how there’s no consistent definition of a planet such that Pluto counts but other small celestial bodies don’t, yet people still mostly consider Pluto the ninth planet? Digital humanities is like that, but with hundreds more Pluto-like appendages. The best definition I can give is when computers collide with any humanities class you took in college. If you really want to learn more, watch the 23 short videos I recorded of Pittsburgh-based digital humanists describing their work a couple years back.
Definitions aside, my day-to-day is surrounded by algorithms. I use them to infer and visualize 17th century British social networks, to understand broad trends in biodiversity research, and to create machine-assisted brief works of fiction, among other tasks. Algorithms are my cyborg prosthesis, augmenting my sight and thought and imagination. They act as macroscope and microscope: lenses that amplify, articulate, and mediate my understanding of society.
As you point out, though, any algorithm that imputes, infers, constructs, aggregates or generally mediates also amplifies its often biased inputs. For example, the Six Degrees of Francis Bacon project I linked above only barely passes the Bechdel test; fewer than 1% of the early modern social ties we inferred are between two women. That’s for two reasons:
We algorithmically inferred social ties from the Oxford Dictionary of National Biography, and most of those biographies are of men.
Our algorithms connected two people whose names appeared together frequently. Because then as now women were often referred to by their relationship with men (the wife/mother/daughter of such-and-such) and their names changed over their lives (taking their husband’s surname), our algorithms often failed to identify and connect them.
We did our best to deal with this, clarifying to readers what was missing, adding relationships from a biographical dictionary of early modern women, and hosting a crowdsourcing women add-a-thon, but that only goes so far. There’s no real way to correct for the fact that the historical record simply biases towards men. We just need to be aware of our sources’ limitations, and take care that our claims take those into account. Historians call this source criticism.
Some people consider algorithmic bias partially a garbage in/garbage out problem. Biased data goes in, biased inferences come out. In the tradition of source criticism, I think of it more as an issue of perspective. Datasets and algorithms come with a particular perspective, and it’s our job to understand the limits that perspective imposes on how we can make claims from or use a particular algorithmic model. Using Microsoft Flight Simulator’s glitchy earth to pretend to fly is lots of fun; using it to navigate the city of Bergen would be idiotic.
My collaborator Chris Warren took what he learned in our Six Degrees of Francis Bacon project to write a biography of the Oxford Dictionary of National Biography. Rather than using an analysis of the dictionary to learn more about British history, he used it to learn more about the dictionary itself, and how gender and class shaped its writing over the years.
One real promise of algorithms isn’t in their direct use, but in exploring their glitches and failures. We can be entertained by Microsoft Flight Simulator’s weird Washington Monument and Janelle Shane’s AI-generated recipes; we can learn more about labor that went into Google’s book scanning initiative by seeing hands that accidentally weren’t algorithmically filtered out, often those of people of color; or we can see the the racial bias in digital image collections revealed by portrait-generating algorithms turning Obama, Oprah, and Laverne Cox white and de-pixelation algorithms turning Obama white.
Algorithmic models hold up a mirror, often revealing ugly truths about ourselves, such as when Microsoft shut down its conversation bot Tay after 16 hours because it began spewing racist and sexist vitriol. Unsurprisingly, the best writers on this subject are women and BIPOC, who tend to be the ones most harmed by uncritical applications of machine learning. Two well-known books on the subject are Safiya Noble's Algorithms of Oppression and Virginia Eubanks' Automating Inequality.
To come back around to your question about avoiding amplifying blind spots, I see part of my job as pointing out those blind spots. We’ll never avoid them entirely, but we can be more careful in circumscribing the sort of claims we make from perspectivally-limited datasets and models. To do that, we need a very careful understanding of how our datasets are constructed and exactly what our algorithms amplify or diminish.
Leah: Scott, I have so many tabs open now! I’m fascinated by the projects you’re highlighting, and, of course, it’s making me think a little about the biases or lacunae in what I read that I’m just discovering them now, when I know every minute detail of particular Internet Dramas that I wish I didn’t. They don’t sound newsworthy per se, but they do sound worthy of my time. And that brings me to another question.
One of the difficulties I’ve run into as a data journalist is that having data can lead you to write about questions that aren’t very interesting or important. It’s a streetlight effect problem. When I give talks to students about data journalism, I show them an embarrassing piece I should have spiked about exclamation point usage by presidential candidates on twitter. It was something I could measure, but I don’t think it illuminated anything or was worth anyone’s time to read.
As you work on research projects, how do you balance a playful, undirected exploration of a new dataset with check-ins that what you’re measuring matters?
Scott: Oh, marvelous question! Digital humanities is to data journalism what a biochemistry department is to a pharmaceutical company, in that we have a little more freedom to explore weird projects that might lead to dead ends in the hopes that some will pay off. That sort of exploration, for example, led researchers to learn that the rate of high-frequency function words like and, of, and the can offer an authorial fingerprint, allowing them to fairly accurately identify authors of anonymous or pseudonymous works.
That said, when you have so much data, it’s easy to lose sight of what isn’t collected or countable.
Mimi Onuoha has a great series of exhibits, the Library of Missing Datasets, highlighting the data to which we don’t have access. Her collection includes underreported incidents of sexual assault and people excluded from public housing because of criminal records. Since we only quantify what we have access to, there are whole worlds of analyses that ought to get published but don’t. To extend your streetlight metaphor, these dark spots are usually (tragically) in places most in need of more light.
Although in research we have a bit more freedom to explore blind alleys, there’s always the concern of going too far and getting lost. If you’ve spent three years collecting and analyzing data on a particular subject that hasn’t borne fruit, rather than giving up all that sunk time cost, you’d probably want to keep reaching until you find something. That’s a common problem in my field.
To counteract that, I practice question- or hypothesis-driven research. Exploratory, data-driven analysis is great when you’re trying to get to know your data, but it can lull us into thinking the data in front of us is all there is, and that it’s inherently useful. It usually isn’t. The “check-ins that what [I’m] measuring matters” ideally come before analysis, when I’m picking research projects and forming questions.
Come to think of it, that’s probably another difference between digital humanities and data journalism—I have the luxury of working on a project for months or even years, whereas data journalists need to churn out more frequent pieces. In a research setting, your presidential exclamation point dead-end might just be a preliminary, unpublished analysis in a larger project on presidential tweeting. I believe Liz Losh has an upcoming book comparing the tweet styles of Obama and Trump.
Maybe the question is less about balancing playfulness and relevance, then, but about ensuring algorithmic analysis has the time and forethought necessary to serve a purpose. Playful exploration is part of the process, but it can’t stop there.
Scott and I will continue our conversation in our final post. Add your comments and thoughts below.