Artificial Intelligence for Cultural and Historical Reasoning is a new international research project funded by Schmidt Sciences that brings humanities and social science expertise directly into how AI tools are built and evaluated.
Language models are increasingly used to analyze texts, generate narratives, and answer complex questions. But most are trained and tested using benchmarks that assume there is a single correct answer.
For researchers working on culture, history, and social change, that approach often falls short.
A newly funded international project, Artificial Intelligence for Cultural and Historical Reasoning, aims to change that. Supported through a major Schmidt Sciences funding announcement that awarded $11 million to 23 interdisciplinary teams worldwide, the project brings together scholars in sociology, history, literature, and information science to develop new ways of testing whether AI systems can reason accurately within specific cultural and historical contexts.
UBC sociologist Laura Nelson is one of the project’s principal investigators, alongside collaborators at Cornell University, McGill, and the University of Illinois Urbana-Champaign.
Working at the intersection of computation and social inquiry, Nelson has played a key role in building research infrastructure at UBC. In 2022, she launched the UBC Centre for Computational Social Science to support interdisciplinary approaches to studying culture, politics, and social change.
In this Q&A, Dr. Nelson spoke about what excites her most about the AI for cultural and historical reasoning project, why historical context matters for AI, and what kinds of research questions these tools could help answer.
What excites you most about this project?
Dr. Laura Nelson, associate professor, Department of Sociology and director of the UBC Centre for Computational Social Science
AI systems, particularly today’s deep-learning models that generate text and images, are usually evaluated using tasks with a single “right” answer, such as standardized tests, math problems, or trivia questions. But social scientists and humanists often ask very different kinds of questions. We study issues that don’t have one correct answer, such as how ideas about gender change over time, or how cultural beliefs differ across societies.
Language models offer powerful new ways to study these kinds of patterns at scale. But they can’t be evaluated using the same benchmarks designed for factual recall or logic puzzles. Testing and developing models for cultural reasoning requires different ways of defining and measuring culture—and that, in turn, requires deep input from the social sciences and humanities.
Doing this well could make AI tools genuinely useful for tackling some of the most challenging questions in those fields.
Why is it important to build models that reflect specific cultural and historical contexts?
If we want to use these models to study culture and history, we need to know they can actually represent different perspectives from different times and places. Just because a model can be prompted to “sound like” someone from Victorian England doesn’t mean it’s accurately reflecting that worldview.
Current models tend to slip into anachronism — including knowledge or language that people in a given period simply wouldn’t have had. If we’re modeling the early twentieth century, for example, the model shouldn’t know about the Cold War or the moon landing. Without that historical grounding, the results can be misleading.
What kinds of questions could these models help researchers explore?
One area I’m especially interested in is how social movements change cultural beliefs. For example, how did societies shift from believing that women should primarily stay in the home to believing that women should work if they want to?
If we had a language model that accurately reflected the early twentieth century, we could experiment with it — for instance, removing key texts from the women’s movement and seeing how the model’s outputs change. That could help us better understand which ideas, books, or arguments were most influential. Rachel Carson’s Silent Spring is often credited with launching the modern environmental movement — but why was it so impactful, and what other works had similar effects?
In a sense, this allows us to experiment on simulated versions of history to better understand the path history actually took.
How does your background as a sociologist shape your approach to this work?
Sociologists spend a lot of time thinking about how to define and measure complex concepts like culture. Culture isn’t just a list of rules; it’s how people make sense of the world, and that’s difficult to operationalize.
We also recognize that multiple perspectives on the same question can be valid. Not every problem has a single correct answer, and that means designing and validating AI models that can account for different viewpoints. That perspective is essential if we want AI tools that genuinely reflect human social and cultural complexity.
