Katherine Elkins studies how increasingly capable AI systems reason, persuade, interpret, coordinate, and interact with people in contexts where judgment, language, history, culture, and institutions matter.
Her current research spans model behavior, human judgment, agentic and multi-agent systems, narrative, behavioral evaluation, culturally and historically grounded AI, and governance based on empirical research.
Current research
The common problem is not simply whether an AI system is accurate. It is how models reason, persuade, interpret, coordinate, and fail when language, judgment, incentives, culture, and institutions shape the task.
Reasoning, ethical judgment, decision-making, persuasion, deception, manipulation, calibration, and linguistic failure modes. This work tests how behavior changes under syntactic reframing, ambiguity, social pressure, and shifts in context—and whether a model’s explanation tracks what it actually does.
Coordination, debate, explanation, adversarial interaction, emergent behavior, and evaluation of agents. Current mentored and collaborative work examines deception in multi-agent negotiation, emotional stability among interacting agents, and the use of agent debate in consequential decisions.
Narrative is a cognitive and emotional tool: a structure for memory and interpretation, a mechanism of persuasion, and a factor in how models and people organize judgment. Current work connects literary study, cognitive science, and philosophy of information to model responses, behavioral effects, human-AI interaction, and vulnerabilities to persuasion.
Current research spotlight
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 research asks how AI systems can retrieve, represent, and interpret cultural and historical materials without stripping away provenance, context, ambiguity, or expert knowledge. Generic retrieval can find a record while missing why it matters, who is authorized to interpret it, or how place and community shape its meaning. Those are system-design problems.
The project joins archival retrieval, knowledge representation, interpretation, domain expertise, and technical infrastructure to make AI systems more culturally and historically accurate. Related work addresses cultural data governance and public knowledge infrastructure, including participation in UNESCO’s AI, IP & Culture Repository co-design process.
Current research
Behavioral evaluation, benchmark design, model comparison, confidence and calibration, red-teaming, ethics-based auditing, and failure-mode analysis. Elkins and Chun conducted the first ethics-based audit of moral reasoning in deployed large language models; current work continues to examine vulnerabilities conventional performance metrics can miss.
Standards, comparative regulation, open-model governance, public-interest AI, and institutional evaluation grounded in evidence about model behavior. Elkins co-leads the MLA team in the NIST CAISI AI Consortium and coauthored the first systematic EU–China–US regulatory comparison after passage of the EU AI Act.
Methodological foundations
How can context, ambiguity, interpretation, and historical meaning remain present in computational research?
The program began with literature and philosophy; its methods now inform research on models, agents, and institutions.
Elkins and 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. Earlier applications had not supplied a systematic procedure for model selection, robustness, smoothing and scale choices, or interpretation across methods.
Developed by Jon Chun, SentimentArcs is a novel ensemble method for comparing narrative sentiment trajectories across texts. The ensemble reduces dependence on any single sentiment system; Middle Reading connects those computational comparisons to interpretation at the scale of passages, scenes, and whole works.
The program includes the first application of explainable AI to narrative analysis, computational comparison of translations, and the first multimodal method for measuring long-form film sentiment-arc coherence across modalities. Each asks what a representation preserves, what it discards, and how those choices affect interpretation.
The same concerns now shape behavioral evaluation, persuasion research, agent analysis, and culturally grounded AI. Can GPT-3 Pass a Writer’s Turing Test?, coauthored by Elkins and Chun, was the first writer’s Turing test of a large language model and made literary imitation and reasoning experimentally comparable.
Selected work
Model behavior & evaluation · 2024
Ethics-based auditing of moral reasoning across deployed language models.
Open models & governance · ICML 2024 oral
A benefit–risk framework for open-source generative AI.
Narrative method · Cambridge University Press, 2022
Rigorous methodology and ensemble comparison for narrative sentiment analysis.
Model writing & reasoning · 2020
An early behavioral test of literary imitation, authorship, and higher-order reasoning.
Cultural and historical AI · Current project
Open AI systems for endangered cultural archives, provenance, retrieval, and interpretation.
Google Scholar → Books and major works → AI in higher education →
Evidence and uptake
The citation record now spans 15 fields and 55 subfields, including AI, digital humanities, literary studies, education, medicine, law, and related fields.