A new project led by a team of researchers from four universities aims to create and evaluate language models that represent past historical periods. The project, "Artificial Intelligence for Cultural and Historical Reasoning," was recently selected for a 2025 Humanities and AI Virtual Institute (HAVI) award from Schmidt Sciences. The $800,000 grant will be split among four institutions: Cornell University, the University of Illinois Urbana-Champaign, The University of British Columbia, and McGill University. Professor Ted Underwood will serve as the principal investigator for the portion of the project at Illinois.
"It's challenging to represent a vanished world, but we hope in doing that, we'll learn lessons that also apply to minority points of view in the present," said Underwood.
A major concern about AI is that language models are too alike, leading to similar responses. This stems from their reliance on massive web-crawled datasets that favor perspectives dominant on the twenty-first-century internet. To address this lack of diversity, researchers need to be able to train models that represent different points of view and social contexts. In the new project, Underwood and his fellow principal investigators will be evaluating and testing training methods so that models can reason and express themselves across cultures.
The team will also build evaluations to assess whether language models can really represent nineteenth- and twentieth-century culture, in both English- and Chinese-speaking contexts. Once the researchers have good benchmarks, they will test different ways of training models that can reason and express themselves like writers in the past.
"When existing language models are asked to represent the past, they often produce caricature and pastiche," said Underwood. "If you ask a model to 'talk like a Victorian,' it will use elaborate syntax and talk about propriety. But that’s not what people in the nineteenth century did, most of the time!"
Principal investigators at partner institutions include David Mimno and Matt Wilkens, Cornell University; Laura K. Nelson, The University of British Columbia; and Andrew Piper, McGill University.
Schmidt Sciences has awarded $11 million to 23 research teams around the world who are exploring new ways to bring artificial intelligence into dialogue with the humanities, from archaeology and art history to literature, linguistics, film studies, and beyond. As part of the HAVI, these interdisciplinary teams will both apply AI to illuminate the human record and draw on humanistic questions, methods, and values to advance how AI is designed and used.
Underwood is a professor in the iSchool and also holds an appointment with the Department of English in the College of Liberal Arts and Sciences. He has authored three books about literary history, most recently Distant Horizons: Digital Evidence and Literary Change (The University of Chicago Press Books, 2019). His current research project involves building language models that incorporate viewpoint diversity and cultural change as central principles. He earned his PhD in English from Cornell University.