I teach the course Qualitative Social Research Methods in the Department of Sociology at the London School of Economics and Political Science, a course that is either a core or semi-core course for students on the MSc Culture and Society, MSc Economy and Society, and MSc Sociology. On this course, students must use one of the standard tools of qualitative research — semi-structured, open-ended interviews, participant observation, or documentary analysis — to reflect on qualitative research in methodological terms. Documentary analysis was added to the course within the last decade and allows us to address Critical Discourse Analysis (van Dijk, 1993) and British Cultural Studies (Hall, 2013; Hall & Stuart, 2021). Working myself in developing Situational Analysis (Clarke et al., 2018) — a methodological extension of Grounded Theory that actively incorporates discourse analysis — I welcome this addition of discursive material to the course.

Increasingly, students are expanding their research material toolkit in this context of discourse analysis to move beyond political speeches, academic texts, newspaper reporting, and television programs. Under documentary analysis, they are studying social media content. This is not surprising, and I welcome this move, spending a lot of time exploring the ethical issues related to social media as data (Crossen-White, 2015; Inckle, 2015; Stein, 2010; Sugiura et al., 2017). I have, however, had some unease about whether and how one can do “good” (Small & Calarco, 2022) research using only qualitative methods with social media. Isn’t some level of quantification necessary?¹ How do I teach a qualitative course to include the kinds of quantification that online data seems to demand? These questions are being pursued within digital methods (e.g., Caliandro et al., 2024; Gray et al., 2020), but I think they need to be taken up by qualitative researchers more generally in the age of AI.

I also believe the ways in which online data defies the distinction between qualitative and quantitative data may say something about the renewed interest in “social physics”. Comte famously defined what we now call the discipline of Sociology as social physics. The “positivist” branches of Sociology have been located in this intellectual history. Florence Chiew (2022) has traced a genealogy of Comte’s social physics in order to problematize the idea of positivism as “bad” while interpretivism is “good”. Chiew reclaims Comte on the basis that he articulated a distinctly interdependent model of the physical and the social — one that Chiew argues aligns with entanglement metaphors associated with Science and Technology Studies. Chiew contends that Comte’s insistence on historical reflexivity is at odds with the characterization of positivism today; Comte was interested in generating laws for social life that were rooted in history. In contrast, the current and renewed interest in “social physics” seems to persist with Comte’s goal to generate laws for social life, but has replaced historical reflexivity with measurement. It looks positivist, a label that I am not attributing any positive or negative values to. The problem is that much of the new social physics, which seems classically “positivist” in its vernaculars and practices, is instead calling itself “qualitative” and “interpretive” so as to avoid the negative evaluations associated with it.

Qualitative, interpretive research has historically defined itself by trading the breadth of quantitative studies (i.e., representativeness) for depth, collecting “rich data” (Charmaz, 2006) through “immersive research” (Matthews, 2025) that aims to provide “thick descriptions” (Geertz, 1973) of social life as opposed to the laws that Comte and positivism sought. Digital ethnography (Hine, 2008) does this in online spaces. But a discourse analysis of digital materials seems to require some type of quantification to do good social science of social media (Caliandro et al., 2024). My own experience working with a large number of academic texts, for example, required quantification to decide which texts should be analyzed closely and carefully using discourse analysis (Nuyts & Friese, 2023).

We live in a world of quantification; online culture and now generative AI are rooted in this quantification and data expansion. To understand this phenomenon, according to the epistemological basis of interpretive research, a lot of qualitative research needs to be mixed with quantitative methodology. I think my students who want to work with social media data require a mixed methods approach. But I also think it is urgent to specify what mixed methods look like in this context of digitization. Historically, mixed methods were careful to conduct the research that goes into different data streams (i.e., the qualitative parts and the quantitative parts), so that the epistemological assumptions of the different data streams were considered. These data streams were then brought together in specific ways. Digitization blurs the relationships between quantitative and qualitative; I worry that specificity is at risk and that this results in poorly designed research. I will use my own expertise in grounded theory to demonstrate this point.

As a bit of background, my PhD supervisor, Adele Clarke, developed Situational Analysis as a methodological extension of Grounded Theory. With it, she formalized her pedagogical practices in teaching Grounded Theory to PhD students in Sociology and Nursing at University of California, San Francisco, on a course she had inherited from her own PhD supervisor, Anselm Strauss. Strauss was one of the two original developers of Grounded Theory, the other developer being Barney Glaser, with their The Discovery of Grounded Theory (Glaser & Strauss, 1967).

For readers who are not familiar with Grounded Theory, one of its hallmark features is its coding practices. Unlike thematic analysis, Grounded Theory describes small speech acts — line-by-line and with a focus on practice and action — in codes that, over time, formulate a description of a key social process that can be compared in developing a theory that is empirically grounded (e.g., Friese & Clarke, 2012). This approach to coding helped frame the infamous debates between Strauss and Glaser in later years. For Strauss, coding was a way of interpreting the social as action, as part of his pragmatist sensibility; for Glaser, coding was a tool for discovering social process that might otherwise have been implicit as part of his positivist sensibility. With Situational Analysis, Adele was pushing on Strauss’s interpretivism, arguing that coding also needs mapping because action and social process are situated (Clarke, 2005). As her students were doing line-by-line coding in her qualitative methods seminar, Adele had a large piece of paper on which she found herself writing down all the elements that were in the situation of action under question — but that the coding process itself was not surfacing. Situational Analysis was developed as an outgrowth of Adele’s teaching, to help researchers interpret how social processes are situated action.

In very different ways — and despite their differences — Barney Glaser, Anselm Strauss and Adele Clarke were all interested in surfacing the implicit. Indeed, one of the key goals of the social sciences more generally has been to make the implicit, or the taken-for-granted elements of social life, explicit (Strathern, 1992). Whether it be the role of comparison in Anthropology’s use of ethnography or Ethnomethodology’s use of breaching experiments, a range of social science methods have aimed to make explicit the otherwise taken-for-granted assumptions that make social worlds work in certain ways. Much of Science and Technology Studies, and particularly Infrastructure Studies, is premised upon extending this focus to material culture and sociotechnical systems (Gray et al., 2020; Plantin et al., 2018). This ability to see the taken-for-granted parts of social life is now considered one of the primary affordances of artificial intelligence for research. What does this shared focus mean? And what differences are obscured? Ethnography and ethnomethodology are, after all, radically different methods for articulating taken-for-granted assumptions, and so we should expect that AI surfaces the implicit in very particular ways as well.

Neither Grounded Theory nor Situational Analysis was developed with the digital in mind. As a result, neither is equipped on its own to contend with the digital as a source of data or, I would add, by extension, as a topic of investigation. Both have been extended in this context, with Computational Grounded Theory by Laura Nelson (2020) and Situational Analytics by Noortje Marres (2020). I want to discuss both, with an aim of seeing what interpretivism looks like when combined with the fact of quantification in digital social life, and to show that mixing can be done in different ways. I ask what is collapsed and what is extended in the process.

I will start with Laura Nelson’s work on Computational Grounded Theory (e.g., 2020), and focus on her 2017 article in Sociological Methods & Research. One of the first things that I noted about this article when reading it was that Nelson does not refer to Grounded Theory coding — what I often think of as one of the hallmarks of Grounded Theory — but rather to Content Analysis (Krippendorff, 2019). Content Analysis is its own research method, rooted in counting content rather than interpreting narratives and discourses. I was confused by this reference, although I quickly came to see that counting is indeed central to Nelson’s approach to Grounded Theory. Nelson states that to “measure meaning” is something that Computational Grounded Theory can do well and that has been central to “inductive Grounded Theory”.

I take Nelson’s term “inductive grounded theory” to be a nod to Glaserian Grounded Theory, although Nelson does put Kathy Charmaz’s Constructivist Grounded Theory into this category as well. This seems strange to me because I do not think that Kathy Charmaz was at all interested in measuring meaning, and her approach to Grounded Theory is different from Glaser’s. It is, admittedly, strategically useful for Nelson to bypass a great number of debates within and about Grounded Theory in order to create a cohesive Computational Grounded Theory. But I would like to resurface some of those debates regarding inductive research and the difference between measurement and interpretive process, in order to think about the possibilities of a renewed “social physics” that aspires to understand social laws, not through historical reflexivity but through measurement. I take Nelson’s Computation Grounded Theory as being part of this renewed interest.

One of the big debates in and on Grounded Theory has revolved around the idea of induction. The idea of building theory inductively, from the data that is collected and produced as part of the research process, became formalized in Grounded Theory. This induction has also been debated and problematized. For example, there was advice in some strands of Grounded Theory for researchers to wait to do the literature review until after the data collection and analysis are complete so as to avoid bias. This sort of research practice is both impractical and inadvisable because a key component of research involves finding out what a group of scholars knows and does not know to ensure one is asking a good and relevant research question. But also — and possibly worse — this advice imagines that the researcher can achieve the status of a tabula rasa and find a “true” social theory as a result. This is simply not possible; the qualitative researcher cannot remove themself from the milieux that they are studying. This ideal — of avoiding researcher bias — represents one of the many positivist assumptions within Grounded Theory that Adele sought to address with her Situational Analysis, as she wanted to replace the idea of researcher bias in Grounded Theory with researcher reflexivity. In this context, much of Grounded Theory has advocated for an approach that is not “inductive” but rather “abductive” — building on the classic idea in Grounded Theory that the researcher should constantly tack back and forth between data collection and data analysis (that includes a literature review and theoretical consideration) through the practice of writing memos.²

What these debates within and over Grounded Theory highlight is the longstanding debate over the meaning and limitations of “inductive” research, which is linked to tensions between positivism and constructivism. Nelson sidesteps these debates, not positioning her Computational Grounded Theory as positivist — despite holding onto some of its vernaculars such as bias, validity and reproducibility — but rather as explicitly interpretive. I think that her use of this phrase risks collapsing together different knowledge practices. The key danger that I see with Computation Grounded Theory, and by extension renewed interest in social physics, is not between qualitative versus quantitative or even big versus small data, but rather the ways in which a neo-positivism is being pursued in the name of interpretivism. I note that qualitative researchers are somewhat responsible for this — often treating positivism as a “bad word”. Neo-positivism is not “bad”, but it is different from interpretivism. It is well established that there are different epistemologies at work in qualitative research (Mason, 2002), and we need to keep some distinctions and specificities in academic language at play — not least to design good research projects.

I turn now to Noortje Marres’ Situational Analytics, which does interpretive research with and of the digital (Marres, 2020). With Situational Analytics, Marres and the research group she worked with analyzed test-drive videos of self-driving cars on YouTube to ask if these videos could be understood as modes of evaluation regarding the introduction of something new into a situation. Marres and the research group did this by first mapping the situational elements in a small subset of online videos. They then conducted a semi-autonomous textual analysis of a much larger group of videos using quantitative methods. What they found is that YouTube’s infrastructure left a mark on their analysis. Situational Analytics does not rely on counting (as content analysis does) nor on uncovering layers of meaning (as thick description in interpretive ethnography does) but on identifying conditions for discourse to emerge.

Situational Analysis focuses on researcher reflexivity rather than researcher bias to explore the situations in which discourses emerge as interpretive research. What Situational Analytics does is expand what reflexivity means in this context. The black boxing of digital systems — whether it be algorithms, supply chains or artificial intelligence — is a key problem and challenge today for researchers and policy makers alike (Gray et al., 2020). Meanwhile, qualitative researchers are using a full range of digital technologies as part of research, ranging from conducting interviews by Zoom, using closed captioning to start transcription and — as my students often do — finding data sources. We can reflect on how these digital technologies — famously opaque infrastructures — shape our analyses with Situational Analytics by asking how computational systems leave marks on our research. It is not enough to note digital infrastructures in our research, as an element in Situational Analysis; we need to use those infrastructures in order to see how they make the situation we are studying what it is.

I think that this matters for the ways in which software for qualitative research is being developed — and here is where the economics comes more fully into view. ATLAS.ti was developed explicitly with Grounded Theory in mind. Adele told me that Strauss was consulted by the early developers of ATLAS.ti, and this can be seen in how it allows the researcher to do line-by-line coding that can then be linked up with a diagramming function of codes as a social process, as well as memos. Grounded Theory was built into ATLAS.ti. In turn, the idea of the social as action was built into the software. My question is which version of Grounded Theory is informing the development of AI in ATLAS.ti? I know from their promotional videos that AI is being incorporated to suggest codes. Has the goal of measuring or interpreting meaning been built into this software? What are the consequences for interpreting data?

Re-emerging discussions of a social physics are opening up new discourses that eclipse important differences, recur older debates, and extend familiar concepts in new ways regarding qualitative and quantitative research as well as positivism and constructivism. I myself would like to see more pluralist ways forward. As a qualitative researcher, I think it is important that positivism is not viewed as a bad word but rather a particular way of knowing that, like all ways of knowing, has certain blind spots and limitations. This might let researchers pursue what looks to be positivist projects without having to do so under the cover of interpretivism. This is necessary to preserve interpretive ways of knowing that are urgently needed, not least because they provide the tools required to explore the kinds of epistemologies that are being built into the digital systems that qualitative researchers use.