The Atlantic By Sara M. Watson
July 1, 2014 11:39 AM
Facebook has always “manipulated” the results shown in its users’ News Feeds by filtering and personalizing for relevance. But this weekend, the social giant seemed to cross a line, when it announced that it engineered emotional responses two years ago in an “emotional contagion” experiment, published in the Proceedings of the National Academy of Sciences (PNAS).
As a society, we haven’t fully established how we ought to think about data science in practice. It’s time to start hashing that out.
Before the Data Was Big…
Data by definition is something that is taken as “given,” but somehow we’ve taken for granted the terms under which we came to agree that fact. Once, the professional practice of “data science” was called business analytics. The field has now rebranded as a science in the context of buzzwordy “Big Data,” but unlike other scientific disciplines, most data scientists don’t work in academia. Instead, they’re employed in commercial or governmental settings.
The Facebook Data Science team is a prototypical data science operation. In the company’s own words, it collects, manages, and analyzes data to “drive informed decisions in areas critical to the success of the company, and conduct social science research of both internal and external interest.” Last year, for example, it studied self-censorship—when users input but do not post status updates. Facebook’s involvement with data research goes beyond its in-house team. The company is actively recruiting social scientists with the promise of conducting research on “recording social interaction in real time as it occurs completely naturally.” So what does it mean for Facebook to have a Core Data Science Team, describing their work—on their own product—as data science?
Contention about just what constitutes science has been around since the start of scientific practice. By claiming that what it does is data science, Facebook benefits from the imprimatur of an established body of knowledge. It looks objective, authoritative, and legitimate, built on the backs of the scientific method and peer review. Publishing in a prestigious journal, Facebook legitimizes its data collection and analysis activities by demonstrating their contribution to scientific discourse as if to say, “this is for the good of society.”
So it may be true that Facebook offers one of the largest samples of social and behavioral data ever compiled, but all of its studies—and this one, on social contagion—only describe things that happen on Facebook. The data is structured by Facebook, entered in a status update field created by Facebook, produced by users of Facebook, analyzed by Facebook researchers, with outputs that will affect Facebook’s future News Feed filters, all to build the business of Facebook. As research, it is an over-determined and completely constructed object of study, and its outputs are not generalizable.
Ultimately, Facebook has only learned something about Facebook.
The Wide World of Corporate Applied Science
For-profit companies have long conducted applied science research. But the reaction to this study seems to suggest there is something materially different in the way we perceive commercial data science research’s impacts. Why is that?
At GE or Boeing, two long-time applied science leaders, the incentives for research scientists are the same as they are for those at Facebook. Employee-scientists at all three companies hope to produce research that directly informs product development and leads to revenue. However, the outcomes of their research are very different. When Boeing does research, it contributes to humanity’s ability to fly. When Facebook does research, it serves its own ideological agenda and perpetuates Facebooky-ness.
Facebook is now more forthright about this. In a response to the recent controversy, Facebook data scientist Adam Kramer wrote, “The goal of all of our research at Facebook is to learn how to provide a better service…We were concerned that exposure to friends’ negativity might lead people to avoid visiting Facebook. We didn’t clearly state our motivations in the paper.”
Facebook’s former head of data science Cameron Marlow offers, “Our goal is not to change the pattern of communication in society. Our goal is to understand it so we can adapt our platform to give people the experience that they want.”
But data scientists don’t just produce knowledge about observable, naturally occurring phenomena; they shape outcomes. A/B testing and routinized experimentation in real time are done on just about every major website in order to optimize for certain desired behaviors and interactions. Google designers infamously tested up to 40 shades of blue. Facebook has already experimented with the effects of social pressure in getting-out-the-vote, raising concerns about selective digital gerrymandering. What might Facebook do with its version of this research? Perhaps it could design the News Feed to show us positive posts from our friends in order to make us happier and encourage us to spend more time on the site? Or might Facebook show us more sad posts, encouraging us to spend more time on the site because we have more to complain about?
Should we think of commercial data science as science? When we conflate the two, we assume companies are accountable for producing generalizable knowledge and we risk according their findings undue weight and authority. Yet when we don’t, we risk absolving practitioners from the rigor and ethical review that grants authority and power to scientific knowledge.
Facebook has published a paper in an attempt to contribute to the larger body of social science knowledge. But researchers today cannot possibly replicate Facebook’s experiment without Facebook’s cooperation. The worst outcome of this debacle would be for Facebook to retreat and avoid further public relations fiascos by keeping all its data science research findings internal. Instead, if companies like Facebook, Google, and Twitter are to support an open stance toward contributing knowledge, we need researchers with non-commercial interests who can run and replicate this research outside of the platform’s influence.
Facebook sees its users not as a population of human subjects, but as a consumer public. Therefore, we—that public and those subjects—must ask the bigger questions. What are the claims that data science makes both in industry and academia? What do they say about the kinds of knowledge that our society values?
We need to be more critical of the production of data science, especially in commercial settings. The firms that use our data have asymmetric power over us. We do them a favor unquestioningly accepting their claims to the prestige, expertise, and authority of science as well.
Ultimately, society’s greatest concerns with science and technology are ethical: Do we accept or reject the means by which knowledge is produced and the ends to which it is applied? It’s a question we ask of nuclear physics, genetic modification—and one we should ask of data science.
By Adam Frank