Sponsored by the Rhodes Information Initiative, the Humanistic AI Conference is designed to foster substantive dialogue between the humanities and the sciences.
The conference will explore questions such as: How might large language models be used to simulate or operationalize historical knowledge? How is generative AI reshaping concepts such as authorship, interpretation, and evidence? And how might current AI models be redesigned to better accommodate humanistic perspectives and theories?
Location: Gross Hall 330 Dates: All day Friday, Sept 25, half day Sept 26 (9-1pm)
Duke University • September 25–26, 2026
Duke Calendar: https://deeptech.duke.edu/event/humanistic-ai-generative-models-and-future-cultural-and-historical-interpretation-09-25-2026/
Speaker Talks and Descriptions
Lauren F. Klein
On Models and Marbles: Reasoning and Interpretation in the Age of AI
Abstract / Description
While reasoning is most often upheld as the goal of AI model development, this talk proposes that a second, equally foundational skill deserves our attention: interpretation. By interpretation, I mean not the domain of logical inference but of meaning-making—the process of asking, given a particular object of culture—a poem, a historical letter, a dance performance, or an Instagram pic–what does it really mean? With examples ranging from the 1972 “Blue Marble” photograph of the earth from outer space, to the “world models” that are currently a focus of AI/ML research, I will argue that reasoning does not merely exist in relation to interpretation, but fundamentally depends on it. I will further argue that expertise in interpretation is what humanities scholars can bring to AI/ML research. Because it is creative, judgment-driven, and irreducibly human, interpretation is precisely what will continue to push forward the bounds of what we can model—and, therefore, what we can know.
Eamon Duede
Abstract / Description
I’ll be talking about intellectual distance and impact in the sciences and the humanities. I’ll be linking results from a study of ~45MM papers to a formal model of decision making to say something about how the humanities could be affected by AI. So, the talk comes from 2 different papers. One written, one not written.
Hoyt Long
Experiments in Mindless Reading with LLMs
Abstract / Description
LLMs are increasingly becoming active cultural agents in human society. They have become meaning-making technologies, for good or for ill, by virtue of their capacity to interact with people through written language and other dominant media. In this presentation, I focus less on the social and cultural implications of their ability to produce written text than on their ability to synthesize and “read” it. I will introduce a series of recent experiments that try to understand whether LLMs can close-read literary texts, what it means for them to read alongside us, and how their specific affordances as readers can best be leveraged for human creativity.
Emily Kadens
Using AI to Automate the Structuring of Historical Data
Abstract / Description
Recent advances in AI-assisted handwritten text recognition now make it possible to digitize historical manuscripts and archival materials at unprecedented scale. Yet transcribed corpora of historical documents cannot be analyzed unless their contents are systematically structured. My project seeks to provide historians with a toolkit to run chatbots that apply a bespoke tagging language to historical documents and output event-focused tagged data, which can be ingested into spreadsheets or a database for querying.
Maria Antoniak
The Garden of Forking Prompts: How Users Explore Narrative Space in Story Generation
Abstract / Description
“The Garden of Forking Prompts” studies how real users iteratively edit their story-generation prompts in the wild, using chatbot logs from WildChat to build two new datasets: WildStories (275,635 story prompts) and WildEdits (24,291 “edit trees” tracing how users branch and revise a base prompt) and a framework of edit types spanning four directions and fourteen story-element targets. We find that revision concentrates on plot and varies sharply by story format, and that explicit prompts drive notably longer editing trajectories—underscoring how static benchmarks miss how people actually co-write fiction with LLMs.
Ted Underwood
ChronoLogic: How Do We Know Whether LLMs Can Represent the Past?
Abstract / Description
I’ll share results from a new benchmark that measures language models’ ability to situate themselves in a range of English-language genres and social contexts 1831-1930. Two take-aways are: that they’ve gotten better at this, and that excluding modern texts from the corpus to prevent hindsight is not sufficient to give them a situated vantage point. It may not even be necessary
