What can digital methods tell us about scientific practice?
Written on July 22nd , 2026 by Wiktor RorotJune 2026 edition of the Society for the Philosophy of Science in Practice Newsletter included my piece on the promises and limitations of digital methodology in the study of scientific practice. It is available here: SPSP June 2026 Newsletter
I was invited by Elis Jones, who guest edited the newsletter—many thanks for the amazing editorial work!
Below is the reprinted paper, with interactive visualizations:
Digital methods have been a major innovation in the recent history of the humanities. In hindsight, it perhaps seems obvious that philosophers would eventually adopt these tools and turn them back onto science itself. But it is not always so obvious exactly how these tools might be useful for philosophers of science, particularly for those (us!) interested in scientific practice.
Since around 2013, when James Overton’s paper on “explanation” popularized the use of these tools in a philosophy of science context, digital philosophy of science has steadily developed as a sub-field in its own right, in part by promising to lift some of the constraints of traditional approaches to philosophy of science. These newer approaches still look to answer central questions of philosophy of science: for instance, how scientific concepts develop and change, what these concepts capture, or how researchers conceive of the aims of their research.
But digital studies can draw, at scale, on a variety of digital resources to answer those questions: metadata and citation networks, digital databases of various disciplines, or, most commonly, large corpora of scientific papers. Relying on computational tools of natural language processing allows researchers to explore and map textual corpora which extend beyond the limits of unaided cognition—what Franco Moretti called “distant reading”. Indeed, such methods can incorporate more text than can be read in a single human lifetime. This can help extend our picture of scientific activities beyond the (sometimes peculiar) favored case studies of our discipline. And although the methods focus on text, they can still inform our understanding of scientific practices—or so I would like to argue here.
Scientific publications are undeniably the primary focus of digital methods. Papers are used to understand what role epistemic concepts like “explanation” play for researchers (aside from the already mentioned work by Overton, more recently this question has been studied by Christophe Malaterre and his collaborators). Digital methods can also answer other questions, like how concepts central to different fields are defined, and how they change meaning and gain or lose prominence (Charles Pence and Arianna Betti focus on such historical studies, whereas in my own work I explore a more synchronous picture).
But what are the actual practices involved in digital methods? An approach called “topic modelling” has particular prominence. Topic modelling assumes that the “topic” to which a given text or its part refers to can be modelled as a probability distribution over words. “Topics” here are best understood as a specific kind of context (as philosophers Jaimie Murdock and Colin Allen notably suggest). If we are talking about the context of deep learning technology, then words like “network”, “transformer”, “supervised” will have a much higher probability than in the context of, say, geology, where words like “drift”, “erosion”, “volcanic” will be more likely.
Topic modelling algorithms, like Latent Dirichlet Allocation (LDA), attempt to infer those probabilities from frequencies of individual words in the empirical sample, and to find terms which differentiate documents within the sample. Words strongly and more uniquely associated with a certain context can then be used as labels for that “topic”. This information can be used as a proxy of what the text is about. There are important caveats though: some parameters, like the number of “topics”, are given by the user, and sometimes the word groupings found via these methods do not correspond clearly to a human-interpretable topic. In such cases interpretation proves particularly crucial— prompting some critics to compare the process to reading from tea leaves, and more generally—highlighting the persistence of a subjective component in these methodologies (probably of no surprise to readers here - more on that below).
As an illustration, you can see, below, an example taken from my PhD dissertation, where I used topic modeling to identify different contexts in which researchers in biology and cognitive science use the concept of ‘communication’ – based on a corpus of over 1.1 million scientific papers across these fields. Individual points correspond to paragraphs mentioning ‘communication’ and they are coloured and clustered according to the “topic” which the model determined–picking out the specific research context in which the central concept is employed, and allowing us to explore how its different uses relate to one another.
There are a few reasons this might all sound familiar: topic modelling is closely related to a broader family of methods inspired by a view called “distributional semantics”. This largely Wittgensteinian approach suggests that the meaning of words is encapsulated in their use, and the use can be approximated by looking at word co-occurrences. This allows encoding meaning as a vector in a multi-dimensional space (called “vector embeddings”). While this might sound a bit abstract, these methods do provide a way of measuring semantic relations. For instance, the difference between vectors for “dog” and “dogs”, and “cat” and “cats” should be identical, capturing the “meaning” of the plural. Similarly, we can expect words like “continent”, “island”, “archipelago” to be more closely related to one another than to words like “deadline”, “submission”, “review”. This approach is also the basis for large language models (LLMs).
LLMs themselves can be useful for digital philosophers too. They can offer sophisticated ways of mathematically representing meaning of sentences and larger documents. A recently proposed alternative to LDA topic modelling, BERTopic, relies on clustering vector embeddings to identify “topics”, whilst methods of “semantic search” (identifying sentences and documents with similar vector embeddings) can be used to identify papers or passages which refer to particular concepts, without relying on keyword searches. The example of my own work, mentioned above, used semantic search methods to uncover individual sentences and paragraphs which refer to the notion of “biological communication”. Such samples can then be used for conceptual analysis, mitigating some of the risks associated with traditional case study use.
But digital methods are not a panacea: they enhance our abilities to study scientific literature, but there are good reasons to worry about this focus on textual outputs of science (not least that various publishing pressures strongly shape the content). That said, publications can function as a strong factor in the reproduction of concepts, theories, and practices. This is well captured within the cognitive approach to philosophy of science, which allows for viewing publications as “cognitive artifacts”, serving a representational function, and affecting researcher’s cognitive performance, including in other dimensions of scientific practice (this view, emerging from the work of Nancy Nersessian and Donald Norman, has been most explicitly formulated as “cognitive metascience” by Marcin Miłkowski).
These methods can also be extended to datasets which relate to practice in different ways, for instance, studies of peer reviews (something recently championed by Marcin Miłkowski). The same methods could be applied to other sorts of “grey” literature: lab notebooks and reports, various tutorials, mailing lists, and so on. These textual records are often subject to different pressures and so may reveal other dimensions of scientific practice.
There are also non-epistemic challenges for digital approaches to philosophy, and the humanities more broadly: notably, the existential risk that humanities at large currently face, with academic funding being restructured to fit into narrowly understood technological interests of some stakeholders. Almost exactly 10 years ago, Daniel Allington, Sarah Brouillete, and David Golumbia published a critique which highlighted how the arguments in favour of digital humanities follow the techno-optimistic rhetoric of Silicon Valley and feed into the “neoliberal takeover of the university”. Indeed, the risks of techno-saviourism have never been more salient than now, when the undemocratic, oligarchic control of increasingly popular technologies impacts not only academic research, but also politics, the environment and many other aspects of our lives. It is worth underscoring here, then, that digital tools do not secure any ideal of “objective knowledge” for philosophy. Study design—as PSP-ers know well—involves a host of (often arbitrary) methodological decisions which shape what quantitative results those tools provide. Those results still need to be submitted to careful interpretation to provide any relevant insights into philosophical questions. In many cases, such interpretative procedures are the only tool for ascertaining the validity of results, as the possibilities for statistical analysis in digital humanities are currently quite limited.
Happily, a vast majority of studies in digital philosophy use mixed-methods or multi- level methodologies, which pair digital, quantitative analyses with a careful close reading and conceptual analysis of the sample that qualitative methods provide. Digital methods, then, are not here to replace the proverbial philosopher’s toolbox, but rather to expand and enhance it.