On LLM Humanities and Social Sciences

To begin with, I firmly believe that the future undoubtedly belongs to what is often referred to as “computational social science.”

Here, “computational” specifically refers to large language models, and “social science” broadly encompasses all disciplines traditionally dedicated to the study of humanity.

As the title suggests: LLM humanities and social sciences.

This is not, nor will it ever be, an independent discipline/subject; rather, it constitutes a “field.” One of the main reasons for my previous confusion and an at-sea situation, I believe, was my failure to distinguish between “subject,” “major,” and “field.” These three categories, defined by increasing degrees of radicalism, are not to be messed up with. It is naïve to hope that a “field” could serve as a “major” for academia-seeking souls.

The paradigm I describe here offers zero contributions when it comes to students’ choices of academic specialization; however, to put it bluntly, it holds “zero” potential for producing publishable research papers—essentially, an Aleph Zero.

Terms like computational linguistics and computational social science, in my view, do not accurately capture the essence of this field. They evoke the image of meticulous and laborious calculation, which may have been appropriate in a previous era, one characterized by manual entropy calculations and statistical machine translation. Today, however, the term “LLM-based” would be more precise.

The current craze for “digital humanities” is characterized by widespread enthusiasm, yet most participants are unclear about what they are actually doing. The primary benefit, however, lies in its apparent legitimacy when applying for research grants. By employing a few data science methods and data processing techniques, and presenting some impressive visualizations, one can label their work as “computational social science” or even establish a “social science laboratory,” which essentially aligns with the concept of digital humanities. I was once swept up in this craze, even considering a friend who expressed reservations as somewhat conservative. Now, I share a more cautious perspective.

After exploring numerous possibilities, especially after reading countless ACL papers often dismissed by the engineering research community as “storytelling,” I have developed a clearer vision of what “interdisciplinary” truly means. Yet, I also question why these endeavors should be the sole domain of computer scientists.

What are computer scientists actually doing here? They are addressing issues like bias in large models’ responses to antisemitism, detecting hate speech on the internet, preserving endangered languages, and interpreting humorous comics. In essence, they are telling stories that belong to linguistics, anthropology, sociology, literature, and the arts. At least within the ACL community, there are few purely computational theories being discussed.

If we consider large language models (LLMs) as entities possessing a basic “sense-making” linguistic capability, then the education/training of individual agents, the collaboration among multiple agents, and the use of these agents to simulate real-world interactions have already transcended the boundaries of computer science and artificial intelligence, extending into the realm of social sciences. Computer scientists have even ventured into “stealing the bread” from the humanities and social sciences.

If social science scholars do not actively engage in this field—claiming or labeling it as their own—the future of the humanities and social sciences will become increasingly precarious, as traditional AI researchers are already leveraging social science concepts to expand their specialized tracks. Conversely, by embracing this logic, which shifts the research focus from humans to LLMs in service of real-world human applications, the possibilities for simulation in communication theory, economic theory, political science, LLM pedagogy, LLM linguistics, and LLM literature become boundless.

In recent papers from Stanford, for instance, we can even evaluate models by “interviewing” large language models; one of THU groups has used LLMs to simulate the game Werewolf, which is based on fuzzy language… Information diffusion, memetics (such as using LLMs to simulate the rise and fall of catchphrases within a small circle), epidemiology—so long as there are people and society, LLMs can be used for social simulation. Although the current “vanilla” LLM capabilities are not yet exceptionally strong, I am confident that with the continuous advancement of supplementary tools and the models themselves, we will eventually achieve our goals.

I acknowledge that this type of research is 60% LLM and 40% social science, posing a significant challenge to social science scholars, students, and institutions. However, reluctance to learn and inability to learn are two different things. Refusal to grant degrees and the inability to confer degrees are two different things. Even if one cannot learn or master it, collaboration remains an option; and unwillingness to collaborate and failing to produce results through collaboration are two different things. I will refrain from naming specific places, universities, or professors.

Let the contradiction continue. I, however, am ready to seize the opportunity and step into the fray.