The digital humanities disappeared right when we needed it most. The last real fight about “DH” happened in 2019. In that year, the literary scholar Nan Z. Da published an audit of “computational literary studies” (CLS) in Critical Inquiry in which she argued that statistical methods for literary interpretation either produced banal results or, if the results were interesting, were not reproducible or robust. This argument, unlike the more fiery political fights of earlier years, did not give rise to another scandal. Da had shown that there was a problem, and she had done so in quantitative terms. After a few responses (of which mine was one), the episode fizzled. DH went more or less silent as the global pandemic occurred.
A year on from the panic-hype cycle induced by ChatGPT, generative artificial intelligence is forcing us to confront a question that CLS and DH have failed to address: What is specific about digital signs? What does it mean that language can run on a computational platform? What does it mean for our understanding of how image and language interact that, using nothing more than an English phrase, you can prompt an AI image generator to create something like DALL-E’s famous astronaut riding on a horse? Cognitive science does not usually attend to such questions, although attention to them is growing.
In a data-saturated, AI-driven world, literary theory and the theory of data have to merge.
Humanists should be best positioned to respond to them, but the DH wars of the early 21st century have left the majority of them with a reinforced hostility to computation in general. A small minority do have the skills to explore our rapidly kaleidoscoping digital environment, but they have not yet produced a research agenda that calibrates theory and empirical inquiry. The claims of literary theory are suddenly subject to experimental hypothesis testing, because generative AI establishes a kind of statistical-textual condition that allows for experimentation on the literary features of language. But we have not yet developed the right questions to put to these systems. The future of DH turns on a form of digital semiotics, on grasping the nature, not just the extent and shape, of digital culture.
The debates that shaped the digital humanities were methodological. They usually came down to what you could find out about literature by reducing it to data and applying algorithms. Da could not reproduce the results of the digital humanists partly because their methods were idiosyncratic and not guided by any clear theory — there was just a sense that the availability of digitized data meant we had to do something new. That sense has grown even stronger with the rise of a genuinely cultural AI that can generate text and image. But DH has retreated into preservation efforts on the one side and “critiques” of AI on the other. Both are valuable; neither answers the call.
While public debates about generative AI tend to focus on questions of machine intelligence and safety, they have largely ignored questions of the concrete meaning of the language and images that these algorithms manipulate, both as enormous data-set inputs and as increasingly widespread outputs. This is where the 2019 confrontation about DH methods can be illuminating.
The virtue of that controversy was that it was a nerdy, technical affair. It mostly took the form of a debate between Da and Ted Underwood, a professor of information sciences and English at the University of Illinois at Urbana-Champaign, making arguments (in these pages and elsewhere) about what statistical science itself is. Da’s contention was that data science must have very strong guardrails: peer-review processes, experimental protocols, and the like. Because texts are giant relational maps, in which words, meanings, and formal effects all interact with dizzying complexity, anything can be quantitatively measured, and spurious significance can be found anywhere. This means that CLS, for Da, relied on intuition and charisma while using numbers to hide that fact. Underwood, for his part, argued that the “boundary between quantitative and qualitative reasoning is growing fuzzier,” and that, while computation had no plausible case to “reveal much about literary judgment” in the past, now “the rules of the game have fundamentally changed.”
It has never been clear what counts as a “significant result” in DH or CLS, a major problem for a field that once hoped to make literary interpretation scientific. The studies Da attempted to reproduce were often unclear about the relationship between the concepts they used — character, genre, plot — and their data equivalents. Partly, this was because virtually all such studies had relied on techniques based on examining the prevalence of pairs of words, which do not stand in any known, stable relationship to genre, character, or plot. AI can now produce, but still not explain, these literary forms. That task should always have fallen to theory.
Finding a way to connect statistical and cultural significance is about more than just the crisis of the humanities. It’s about our more general, global crisis of digital data.
CLS proponents had sometimes argued that they were replacing “charisma” in literary interpretation with “science,” but Da’s study showed how deeply compromised that idea was. This problem is not specific to literary studies. In fact, for virtually all computational science, the relationship of data to meaning is completely unclear.
Da’s paper was a field audit, an attempt to see if results stood up. The genre has become important as an ever-widening “replication crisis” hits the social (and many other) sciences. As fundamental results fail to reproduce, an enormous problem of significance has emerged across a wide variety of disciplines. The psychologist Gerd Gigerenzer has argued that this is a consequence of the automation of judgments of statistical significance replacing discrimination with data manipulation. The outcome is that results are built on prior results that are not themselves secure, spreading rot across the entire edifice of knowledge-production. Against this backdrop, we can see that Da was including DH in the bigger knowledge crisis that a data-heavy society faces.
Today’s AI feeds off the big data that creates that same crisis. without which its learning algorithms would not work. This is what invests AI with its millenarian hopes: It will solve climate change, poverty, and more, we hear. In reality, the suite of software tools that AI puts out create a de facto interpretation of the very data AI relies on, thereby further corrupting the edifice of knowledge production without enlightening us.
AI depends on the surfeit of data that has caused the significance crisis, a situation that Da’s and Underwood’s positions together reflect. As Underwood and his co-authors put it recently, “It is admittedly alarming that we are now all part of an experiment where theories of language and culture get tested on live subjects.” We need to move past the defensive posture of alarm. Only a theory of culture that is sophisticated about computation has any hope of cutting through the significance crisis, the toxic mixture of data and algorithm that has so watered down the knowledge project today.
To date, and in spite of an early, widely read call from the digital humanist Alan Liu, DH has contributed virtually nothing to literary theory. Nor has it produced a working set of theories that guide the process of data analysis and restrict what can count as a significant result — something that data science itself, of course, does. In a data-saturated, AI-driven world, literary theory and the theory of data have to merge.
Underwood recently called the emergence of generative AI “the empirical triumph of theory.” “Ferdinand de Saussure’s distinction between parole and langue,” he wrote, “is concretely dramatized every time a user sends a prompt to a model.” Literary theorists have long argued that language was not a tool but a medium, not just something that minds manipulate, but something that plays an active role in the formation of knowledge. The name for that medium is not just language but “culture,” and now we are seeing even cognitive scientists forced to admit that one of their greatest achievements to date — generative AI — is a culture machine. Questions about the structure of language, and problems of significance and meaning in culture, are on display in data-heavy learning algorithms. Whether we take empirical advantage of this new availability depends on theory.
The encounter between digital technologies and the humanities has entered a new phase. The first two decades of the millennium perseverated about the proliferation of data, which the word “big” failed to capture. We have now rightly turned to the question of what and who parses data, and how. And the buzz about generative AI can be seen as a belated recognition that the manipulation of semantic data — data that carries meaning, as broken-down words or images, among other things — both produces culture and promises to interpret it. As much as I am on record agreeing with Da, I am less certain that the bright line she draws between interpretation and data-parsing holds at the industrial scale. What is called “machine interpretation” is too common and influential to be dismissed entirely. But this “interpretation” — which results in the recommendations for music, TV, and restaurants that you encounter every day online, for example — is itself poorly understood. The significance of the digital remains an open question.
Establishing formal and rigorous definitions of culture, language, and interpretation in light of but not dependent on algorithms should be the goal of the next phase of the digital humanities. All other paths forward strike me as ancillary to the toxic cocktail of digital capitalism and the replication crisis. Finding a way to connect statistical and cultural significance is about more than just the crisis of the humanities. It’s about our more general, global crisis of digital data.
Much great humanist thinking has emerged from a genuine interest in the sciences. Kant’s philosophy is, among other things, a reaction to the picture of the universe that Newton had left behind. Wittgenstein and his acolytes in the Vienna School — so important for Anglo-American philosophy after them — were deeply influenced by Einstein, Mach, and the early discoveries of quantum phenomena. Neither they nor Kant suggested a “physical humanities,” but instead wrote philosophies that absorbed and gave human context to the scientific world picture.
Structuralism and poststructuralism brushed against cybernetics and information theory, as historians have meticulously documented. Roman Jakobson, Claude Lévi-Strauss, and Jacques Lacan all enthusiastically followed the rise of these sciences, which laid the groundwork for today’s AI. But the synthesis of literary theory and information theory was abandoned and remains a torso. If Underwood is right that language models make that theory empirical, the time for its completion is now. That completion runs directly through data culture and the AI that is quickly becoming its infrastructure. The history of computing, rooted as it is in forcing physical systems to perform representation, can be seen as a progressive simulation and extension of cultural systems. The humanities, whether we call them “digital” or not, are the only division with the potential to make sense of those systems. It seems possible that that potential is precisely what has brought them under fire today.