About
Katherine Elkins
Katherine Elkins is an AI researcher studying model behavior, human judgment, agentic systems, narrative as a cognitive and emotional tool, and how to build AI systems that are more culturally and historically accurate.
Her work combines computational methods with expertise in language, narrative, interpretation, history, and human reasoning to study how models reason, persuade, coordinate, fail, and interact with people and institutions.
Current research
Models, agents, people, and institutions
Model behavior and judgment
Elkins works on technical research questions in model evaluation, behavioral testing, reasoning, calibration, persuasion, deception, linguistic vulnerabilities, and robustness. The work asks what models actually do when prompts, syntax, social context, or incentives change—and when their reported explanations do not match their behavior.
She also studies agentic and multi-agent systems: coordination, debate, interaction among models, and the emotional and strategic dynamics that emerge when agents negotiate or make decisions. A Notre Dame–IBM Technology Ethics Lab project compares human, statistical, and language-model judgments in high-stakes predictions of recidivism.
Elkins co-leads the five-member Modern Language Association team in the NIST CAISI AI Consortium, connecting language, framing, interpretation, and human judgment to federal work on AI measurement and evaluation.
Current research
Narrative, cognition, and emotion
Narrative and cognition
Narrative remains an active research object in Elkins’s work. She studies it as a cognitive and emotional tool: a structure for memory, a mechanism of persuasion, a way people organize judgment, and a means by which models and humans interpret information.
These questions connect current work on model behavior to earlier computational studies of emotional arcs, literary imitation, translation, and generated language. Narrative can support understanding, but it can also create behavioral effects and vulnerabilities. Systems able to model narrative expectations may reason about people more effectively; the same capacities can shape persuasion, manipulation, and trust.
Cultural and historical AI
Archival Intelligence
Rescuing endangered cultural archives.
Elkins is Co-PI of Archival Intelligence, a Schmidt Sciences Humanities and AI Virtual Institute project building open AI tools for endangered archives, beginning in New Orleans.
The project asks how AI systems can work with cultural and historical materials without stripping away provenance, context, ambiguity, or expert knowledge. Its research joins archival retrieval and interpretation with domain expertise and technical infrastructure, with the goal of making AI systems more culturally and historically accurate.
The work extends to questions of cultural memory, intellectual property, creators, and digital cultural sovereignty, including Elkins’s contribution to the UNESCO AI, IP & Culture Repository co-design process.
Methodological foundations
Interpretation made testable
Methodological foundations
A continuous question runs through the research: how can difficult human phenomena be made empirically testable without reducing away context, ambiguity, interpretation, history, or meaning?
Elkins and Jon Chun introduced the first rigorous methodology for narrative sentiment analysis, establishing a reproducible framework for selecting, comparing, validating, and interpreting sentiment models across narrative texts. The Shapes of Stories, Middle Reading, and the novel SentimentArcs ensemble method developed ways to compare emotional trajectories while retaining interpretive accountability.
That work led into the first writer’s Turing test of a large language model, the first application of explainable AI to narrative analysis, computational comparison of translations, and multimodal research on long-form film sentiment-arc coherence. The same methodological concerns now inform evaluation of model behavior, persuasion, judgment, agents, and culturally grounded AI.
Literature and philosophy remain intellectual resources here: they supply difficult cases and developed accounts of consciousness, memory, perception, narrative, and interpretation. They are foundations for the technical questions, not a separate professional identity.
Teaching and field-building
Students as researchers
Teaching and field-building
At Kenyon, Elkins and Chun began developing what Kenyon describes as the world’s first Human-Centered AI curriculum and lab in 2016. The model gives humanities and social-science students the technical fluency to formulate and conduct original computational and AI research while keeping domain expertise central.
Across more than 190 mentored projects, students audit models, build benchmarks, study agentic and multi-agent systems, test emerging capabilities, and develop computational methods. The aim is participation in research, not generic AI literacy.
Current roles
Selected roles and affiliations
- NIST CAISI AI Consortium
- Co-lead, five-member team representing the Modern Language Association; 2024–present.
- Archival Intelligence
- Co-Principal Investigator, Schmidt Sciences Humanities and AI Virtual Institute.
- Kenyon College
- Professor of Comparative Literature and Humanities; Co-Founder, Human-Centered AI Lab.
- Research profiles
- Google Scholar · ORCID · CV