Ethics as Methods: Doing Ethics in the Era of Big Data Research—Introduction

Ethics, Methods, and Accountability

As Markham (this issue) summarizes in the afterword of this Special Issue, it is impossible to standardize or universalize what constitutes the ethically correct actions in technology design and research contexts, not least because we cannot predict what will happen as a result of our choices. Obviously, a belief in and understanding of core human rights, plus basic human decency, plus common sense, can help us make good decisions to minimize the potential harmful impact of our actions to persons or society. But our choices are never simply a matter of using our personal calculus: regulations, laws, and policies governing research also define ethics for us. They do so in necessarily generalized terms to provide broad scope and applicability. More informal structural norms we encounter in our professional practice may prescribe particular additional guidelines. So even though ethical decision making might be ideally grounded and produced in practice, through iterative judgments in specific cases, they are presently locked in the structure side of the structure/agency, or structuration cycle (Giddens, 1979). Or, to employ a Foucauldian framing, ethical guidelines become hardened into incorrigible and obdurate governmentalities (Dean, 1999; Foucault, 1977-1978/2007). This is not an easy situation to unpack, since it is associated with over-regulation in some research-related domains and under-regulation in others.

Meanwhile, the cases of dubious or thoughtless ethical behavior multiply, creating a worrisome ethics crisis. Major elements in this tangle include an “anything-goes” startup attitude that pushes technology designs into the market without adequate scrutiny for their potential impacts, glacially slow updates to conceptual and regulatory guidelines, a push for ever-more precise profiling of people at any cost, and a continued rapid pace of technological transformations that continually change the rules of the game (Markham 2015, Markham and Buchanan 2015).

Alongside this broad crisis, we can identify an equally important failure in the academy. Many, if not all, systems of research and ethics governance do not facilitate and support social research that needs to be done, because it has been traditionally been prohibited. For example, research on children is woefully sparse because children are improperly classified as a single standardized category of vulnerable subjects. As another example, corporate researchers, academics, journalists, and engineers all have different ethical norms governing and guiding their research even as their domains of study overlap. Our understanding of what ethics means and how it might be best enacted therefore suffers. It is our position that embracing a methodological framework of ethics enables productive discussions across domains.

At the level of practice, defining ethics as method draws attention to the epistemic rather than the ontological, or how ethics are enacted, rather than what they are. If taken seriously, this stance brings the concept of ethics from an easily reified abstraction to an always-already emergent and deeply contextual personal practice: anyone doing inquiry of any sort can reflect on how they are personally performing the means. This becomes a necessarily case-by-case process. As Markham writes in 2006, “reflexively interrogating one’s methods of inquiry shifts attention away from codes of conduct imposed from the outside and reveals hidden ethical practices from the inside” (Markham, 2006, p. 39). It also may be used to reflect on how ethics are built not just through the paths that are created when we make a decision to go one way or another at a critical juncture but also in the paths that are closed off, unfollowed, or neglected.

In broad strokes, this approach blends situational ethics and a feminist ethic of care in ways that enact what Judith Simon (2015) discusses as “distributed epistemic responsibility.” As a paradigm, situational ethics builds on the work of the Episcopal priest Joseph Fletcher. Fletcher wrote in 1966 that all choices should be based on the context and circumstance of their particular situation, and not on some universal law. He listed “Love” as the only exception to this situationalism—or rather than an exception, a key mind-set that fosters “good” choices. Although matters of the heart may appear incidentally in contemporary discussions of research ethics, they constitute a core feature of Markham’s (2006) ideas about how ethics are, or should be, enacted as method. “Heart,” she explains, is “an amalgam of consciousness, mindfulness, honesty, and sensitivity,” and an ethical researcher is one who “works from the center,” which entails “being knowledgeable and prepared; present and aware; adaptive and context sensitive; and honest or mindful” (p. 44).

This stance is a hallmark of a feminist ethic of care, yet as defined by Markham here, it seems less about love, specifically, and more about readiness, which draws on Allport’s (1935) definition of the sort of preparation or attitude that is essential to make “satisfactory observation, pass suitable judgment, or make any but the most primitive type of reflex response” (p. 806). In the entangled contexts of data-driven research, however, the consciousness and adaptability that come from readiness may seem ineffective when the individual researcher might play only a minor role—working in large teams, encountering only whatever has been scraped and filtered by an API, or studying data long removed from its human origins. We argue that despite the challenges posed by automated systems, returning to a baseline definition of method as a series of axiological choices can help us move toward more appropriate understandings of how epistemological responsibility (Simon, 2015) moves from being assigned (or distributed) to being enacted, in that we turn our focus toward the specificities of the actions that constitute choices that matter.

Methods as Ethics: How the Tools Under Our Methods Produce Ethics

The surface level accuracy of big data analytics implicitly valorizes and explicitly fosters an old school correspondence theory of truth. The concept of “data” presumes the possibility that human behavior can be traced, isolated, collected as data points, and measured. Big data, as a process of aggregating and performing calculations on diverse and huge sets of data points at high speed, yields promising “truths” about who we are and what we therefore want. The tools used to yield these findings are at some level making interpretations about what the data mean, but the interpretive stages occur early in the processes of testing so that by the time the data analysis yields an effective result, it is highly rule based and computational. Thus, even as we see a renaissance of nuanced interpretivism in the study of humans and society, the discourses around data science promote, whether deliberately or not, a return to the most conservative sorts of positivism.

Along with the authors of this Special Issue, we define and confront this as an emerging grand narrative. The basic premise is that knowledge derived from data analytics is true (or has strong truth value) because of the objective qualities of data, their means of collection and analysis, and the sheer size of the data set. This deceptive premise has been repeated in many forms: in the late 1990s, distance education was heralded as a way to “deliver” knowledge, as if the essence of knowledge was embedded in the content of the material transmitted online to the learner. Now, platform and service providers assert that their algorithms are impartial, functioning through rule-based calculations on objective data points without any “subjective” interference. This standardization supposedly secures their position as “legitimate brokers of relevant knowledge” (Gillespie, 2012, p. 180). The rule-based process—or as we emphasize, the method—of the algorithm indeed yields almost divine-quality results, especially when we see these outcomes as eerily-appropriate advertising on our own screens. The by-product of common discourse around big data not only diminishes or removes the qualitative aspects of behavior and experience that form the data in the first place but also removes the human from the process of analysis. This situates data science as a process of analysis performed by the tool, which removes decision making and judgments from the endeavor.

Historically, the early 19th century science of craniometry is a good example of how our tools both use and produce an ethic, or ethos, which carries significant axiological weight in the design and conduct of scientific inquiry. In the scientific tradition of craniometry, skull size was associated with intelligence, and it just so happened that Caucasians had larger skulls than people categorized into other races. The choice to define intelligence as a solid and measurable variable can be seen as a methods move that, over time and restatement, creates plausibility. We have rejected measuring skull size as a valid method of determining intelligence, but the ethos of reducing intelligence to a numeric value (e.g., bell curves, IQ test scores) remains. This can be seen as an ethics move, in that the use and valuation of particular methods produce or reinforce particular ways of knowing.

The authors in this Special Issue draw out contemporary corollaries. Relevancy, for example, is used in big data analytics as if it is not a subjective category but a natural choice for how search engines should yield results for users. But as computer scientist Joanna Bryson and her colleagues found, algorithms draw on the same biases that humans have (Caliskan, Bryson, & Narayanan, 2017). In this and other cases, the ethos, it turns out, disenfranchises already-marginalized groups (cf. Eubanks, 2018; Noble, 2018). Therefore, algorithms enact a different form of an-ism. As Karppi (this issue) points out algorithms that use relevancy in predictive policing make distinctions and mark “hot spots,” which shapes how people are treated by authority figures. This in turn can influence the resources certain people have access to, how likely they are to succeed at various endeavors, or how likely they are to be penalized by governmental agencies (see also Brayne, 2017).

Often, the ethic is engendered not directly through the actions of the researcher, but indirectly through the absence of questioning the validity of variables in a world that has long since discovered (Kuhn, 1962) that our basic paradigms about what things are, or how they work, are not naturally “true,” but an outcome of debate, persuasion, and other social interactions among scientists. If we apply some of this critical reflexivity toward tools such as APIs, which extract certain information as data while ignoring other information as non-relevant, we can begin to move beyond the obfuscating critique that they operate as black boxes, and begin to question what they are actually producing.

As Rob Kitchin (2014) notes, social research methods have traditionally had to “extract insights from scarce, static, clean and poorly relational data sets,” whereas in the era of “data-driven” research, the challenge is to cope “with abundance, exhaustivity and variety, timeliness and dynamism, messiness and uncertainty, high relationality, and the fact that much of what is generated has no specific question in mind or is a by-product of another activity” (p. 2). We use machine learning systems to cope with this messy abundance, since the size of any single data set and the aggregated complexity of multiple data sets are beyond the processing capacity of human cognition. However, machine learning is becoming “a methodological substrate for knowledge and action . . . a kind of dark matter that invisibly distorts the distribution of benefits and harm” (McQuillan, In Press, p. XX). And, while it is, by now, a well-known and perhaps battered point that datafication has transformed how social research is conducted, more subtly, it impacts what we think we are doing when we conduct said research. A dangerous erosion of the role and meaning of interpretation seems to accompany the shifts that “data-driven” research and interventions have introduced (Markham, 2017).

While this epistemic shift toward “algorithmic knowledge production” (Metcalf et al., 2016, p. 6) highlights the differences and tensions between what we have conventionally considered good social research and what we think big data research is, efforts are being made to synergize the overlaps between the two. Pink and Lanzeni (this issue) suggest that future-focused anthropology and big data research have some things in common. Both share improvisational data gathering methods, encounter the world as it unfolds, focus on the futures, and have interventional ambitions. Somewhat similarly, Halford and Savage (2017) suggest “symphonic social science” as a way forward. Symphonic social science can be achieved through compiling multiple different data sources into a single, unified, and often repeated “refrain,” which in turn makes the arguments created via such methods more persuasive (Halford & Savage, 2017). The similarities (multiple data sources, emphasis on correlation, and use of visualization), and differences (i.e., symphonic SR includes rich theoretical awareness, carefully chosen data, focus on long-term trends) between symphonic social science and big data research are what Halford and Savage (2017) think could extend both and offer a methodological solution to the epistemic shift underfoot.

Agency, Responsibility, and Morality

If we agree that choices (in research) should be contextual and situational and that moral situations involve a relationship between, at least, an originator of an action—or a moral agent and the recipient of actions—or a moral patient, then the actions of the agent and the harms or benefits to the patient can and should be evaluated (Floridi & Sanders, 2004). This formula seems reasonable: It implies that the ethicality of any situation can be assessed either from the perspective of the agent (i.e., responsibility based evaluation) or from the perspective of the patient (i.e., rights based evaluation). But it offers a too-simple binary in the context of datafication, digitalization, and automation.

Perceived agency for ethical decision making has become complicated in the context of machine learning. Big data, which Kitchin (2014) describes as voluminous, swift, diverse, exhaustive, fine grained, relational, and flexible, comes with a “class of machine actions, where the traditional ways of responsibility ascription are not compatible with our sense of justice and the moral framework of society” (Matthias, 2004, p. 177). How do we attribute moral agency, moral patiency, or shared responsibility when neural networks are “obfuscated by nature” (McQuillan, this issue), or when algorithms function as cultural techniques that seem to have their own agency (Karppi, this issue)? At the same time, should we accept the idea that we no longer have agency? Magalhães (this issue) contends that despite our inability to comprehend algorithms, human actors, even non-expert end users, are still capable of perceiving what algorithms do—an insight that grants people some moral power and agency.

We can call this complication of locating moral agency and responsibility a wicked problem. There are no straightforward boundaries, definitions, or answers. Rather, there are only questions to be continually addressed. As Gunkel (2017) notes, almost drily

how we decide to respond to the . . . machine question will have a profound effect on the way we conceptualize our place in the world, who we decide to include in the community of moral subjects, and what we exclude from such consideration and why. (p. 245)

The dilemmas are not new, of course. Discussions of machine ethics often reference the five laws of robotics that the science fiction writer Isaac Asimov posited already in the early 20th century. The first industrial revolution is peppered with machines that critics denounced for dehumanizing worker and separating the mind from the hand, conception of the purpose of work from its execution by labor (Braverman, 1974). Shelley’s Frankenstein yields a different sort of speculation about non-human morality. Yet, by and large, human exceptionalism still stands as the dominant ethical paradigm, and it positions technology as a tool created and used by human beings. This presumption has significant limitations in the light of recent technological innovations (Gunkel, 2017). Google’s go-, shogi-, and chess playing AlphaZero, their facial recognition deep learning network FaceNet, Microsoft’s hate-spewing chatbot Tay.ai were all designed to create their own instructions and evolve their behavior. The people who created them thus have limited control over what the systems do. If neural and deep learning networks develop “minds” of their own, but we view ethics as an issue of agency, then what happens to our ethics when agency blurs? At what point can we hold an API or an algorithm accountable for its decisions and actions? Or how can we hope to control them (cf. Karppi, this issue; McQuillan, this issue)?

While there is no consensus on whether we can assign moral agency or moral patiency to machines and systems, “what is not debated is the fact that the rules of the game have changed significantly” (Gunkel, 2017, p. 237). Even if we do not think that machines can or should have a moral status, it seems that the ability to act has escaped human confines. The time to ask ourselves whether it is ethical to build systems that have the ability to act without having the ability to take on moral agency, too, has passed, as we have already built them.

Returning to the question of how we can identify morality in these entanglements of agency, we agree with Silverstone (2006) that it is far more important to consider matters of responsibility and accountability. These concepts call our attention to the idea that someone needs to take responsibility for responding, when something goes awry. This shifts the focus in a pragmatic and future-oriented way.

The Role of the Researcher

If “unraveling the intricate tapestry of method and ethic . . . involves partitioning what appears to be a smooth flow of one’s choices and movements during the entire research project” (Markham, 2006, p. 39) then how do we accomplish that in the context where automated systems play an increasing role in those movements, and possibly alter what we experience as choices? Most of us can probably agree that good researchers strive to find situated, context-appropriate ways for linking their habitual decision making with morality. But there is less consensus on how to actually accomplish this. We suggest it is a matter of interrogating what is happening along the string of actions that eventually lead to the construction of data, focusing on human practices and nonhuman processes that function interpretively, and reflecting on various levels and stages of possible impact of such actions, constructions, and interpretations.

Together, the collaborators of this Special Issue coalesce around two broad suggestions for building more reflexive practice:

For example, Zimmer (this issue) offers an ethics assessment heuristic based on Nissenbaum’s (2004, 2010) “privacy as contextual integrity” framework. He points out that this heuristic can empower researchers to be “more attentive of the normative bounds of how information flows within specific contexts” (Zimmer, in Press). To demonstrate his chosen heuristic in action, Zimmer (this issue) applies its nine steps to the much-critiqued 2016 case, where a Danish graduate student released a large amount of OKCupid user data to the public with the claim that it was “for science” and OK, because “the data was already public.” Zimmer persuasively shows how much harder it would have been to dismiss the moral and ethical problems of the “already public” claims, if the data scrapers and sharers would have assessed its contextual integrity and not just its “public” accessibility.

A different heuristic is offered by Luka and Millette (this issue). They describe speculation as “a gathering of possibilities” that allows researchers to practice care by making their own voice one among many, becoming aware of, and sensitive to, their own positions, biases, and power. They queer Hannah Arendt’s (1961) concept of action ¹ to analyze the “unacknowledged costs involved in data production and analysis.” This might be enacted through processes of superseding, supplanting, or augmenting big data by small, lively, or thick data (or the other way around) within the same project (cf. also Latzko-Toth, Bonneau, & Millette, 2017). Whereas Zimmer’s approach relies on the framework of “contextual integrity” to be able to take on an alternative perspective in regard to the research project, Luka and Millette suggest designing studies with, rather than upon others, to create a context-specific heuristic:

Sharing experiences from their own sensitive research project with men who have sex with men, Light et al. (this issue) show how a particular aspect or a “data point” can become particularly sensitive in specific contexts and situations. They offer an example wherein location becomes such a key factor for big data ethics. Operationalized through stages of gathering, analyzing, and presenting, as well as archiving and deleting data, the authors show how the most ethical choices may actually be those that are methodologically counterintuitive. They emphasize that not all that can be collected should be; that data visualization has its own politics, and thus ample ethical problems; and that even though storing data and sharing code is lauded in social research as a tool of increased validity and reliability, an ideologically communal act, or a way to get more “bang for public buck,” the most ethical choice may be to delete both data and scripts.

Sharing her insights from studying extremism—arguably the current plague worldwide, and an issue undoubtedly entangled with big data, algorithmic sociality, and social media—Massanari (this issue) highlights the need to mitigate researcher risk under these new conditions. However, as she emphasizes, the need to reduce harassment and intimidation of academics clashes with the dominant norm of academic visibility and microcelebrity. From this junction, Massanari provides concrete suggestions for professional practices that work well to support researchers, share best practices, and build stronger and broader understandings of risk.

These two broad suggestions may resonate with our own attitudes and inspire us. Yet it is also quite clear that they do not easily link with or accomplish formal requirements of funding agencies, PhD programs, and ethics oversight institutions. This can create some cognitive dissonance for researchers seeking to do what is actually ethical, when it may not align with extant regulatory guidelines (Whelan, this issue). This dissonance is likely quite common at present, as it is well understood that “insofar as research ethics regulations and conceptual frameworks are responding to the conditions of knowledge production that precede the epistemic shift towards algorithmic knowledge, (big data) research and the extant research ethics regimes will be mismatched” (Metcalf et al., 2016, p. 6).

There are new models to help ease such dissonance and link methods to ethics and vice versa. Two influential groups are worth mentioning for their steady insistence that as scholarly communities, we owe it to ourselves, future generations of scholars and humans, and perhaps, idealistically, the state of the world, to push against administrative, corporate, regulatory, and legislative frameworks that fail to accommodate the needs of contemporary ethical practice in digital media use and social research: The ethics working group of the Association of Internet Researchers (AoIR) ² , who are now working on their third iteration of international guidelines for ethical decision making in social research, and American University’s Center for Media and Social Impact ³ , who have for many years developed industry-specific codes of best practice that challenge outdated interpretations of copyright.

Our special issue joins this growing number of voices imagining, experimenting with, and sharing the practices of enacting ethics as methods and methods as ethics. The authors all agree in different ways that we cannot afford to be paralyzed by the wicked problems and the uncomfortable researcher subjectivities prompted by technological, political, economic, and institutional complexities. We hope the contributions in this Special Issue inspire readers to challenge how ethics are conceptualized within ill-fitting regulatory and legal frameworks, explore how our tools and techniques carry and produce an ethos that over time becomes taken for granted, and operationalize epistemological concepts that fit the complexities of agency and accountability in data-entangled research contexts.