Scholarly reception & cross-disciplinary influence

Jon Chun’s independently authored SentimentArcs developed an ensemble method for comparing sentiment trajectories across long-form narratives. The method combines model comparison with time-series techniques, including dynamic time warping, that make narratives of unequal length computationally comparable.

Later researchers have used the approach across several kinds of narrative data.

The method has been applied to novels and fan fiction. Other work has used related trajectory analysis for games and screenplays. Researchers have also taken the approach into end-of-life medical narratives and economic-crisis discourse.

The open-source implementation has been maintained since 2019 and became the technical foundation for The Shapes of Stories.

SentimentArcs also established a technical line that later work extended. Chun and Elkins added sentence-level explainability and shared-coordinate comparison in their 2023 XAI paper. Chun’s MultiSentimentArcs then carried trajectory comparison across modalities by testing whether dialogue and image produce the same emotional arc in film.

The technical work grew out of earlier research with Katherine Elkins on nonlinear narrative and “middle reading,” a method positioned between distant reading and close interpretation.

Later researchers returned to the approach in work on Hemingway and Danish literature. Other scholars adopted its smoothing methods or cited “middle reading” directly as a digital-humanities methodology.

The importance of this work for Jon’s later research is methodological: it established a way to turn difficult interpretive questions into measurable problems without treating interpretation as noise.

Chun and Elkins developed an explainable-AI workflow for long-form narrative analysis that linked model outputs back to sentence-level evidence and addressed the problem of comparing narratives of unequal length.

Later work by Cugurullo and Xu cited the approach in Policy and Society. Surveys in IEEE Access and Discover Applied Sciences included it in broader accounts of explainable generative AI. Other researchers used the approach in LREC workshop research and in applied banking work.

The reception is significant because the method moved beyond literary analysis into policy and applied AI research.

Chun’s independently authored MultiSentimentArcs extended narrative trajectory analysis from text to film.

The method asks whether the emotional trajectory recovered from dialogue and the trajectory recovered from images form the same arc. This turns cross-modal narrative coherence into a measurable quantity rather than an interpretive assertion.

Chun and Elkins’s 2020 writer’s Turing test treated literary imitation as an empirical test of GPT-3.

Floridi and Chiriatti cited the experiment within a year. Later researchers returned to it in work on machine psychology, impersonation, literary memorization, authorship, and narrative bias.

The paper also informed later work on human and machine persuasion, including research published in Science Advances and PNAS.

Chun and Elkins developed an ethics-based audit of moral reasoning across deployed large language models that included confidence scoring.

Later researchers adopted parts of the experimental design and carried them into new evaluations.

Liu and colleagues used the confidence-scoring method for value-priority evaluation at COLING 2025. Sowmya and Vasudeva replicated the eight-model audit in IEEE Access. The Edinburgh LLM Ethics Whitepaper included the work in its discussion of methods for evaluating model values and ethical judgment.

The work forms part of the methodological background to Chun’s current research on model judgment, explanation, and evaluation under ambiguity.

The EU–China–US framework developed by Chun, Schroeder de Witt, and Elkins has become a widely used reference for comparing national approaches to AI regulation.

The work appears in research published in Communications of the ACM, PNAS Nexus, Nature Communications, Information Fusion, and comparative-law scholarship.

More recent work uses the framework more directly. Prabhakar and colleagues build an evaluation framework for national AI regulation on the three-regime comparison. Malanond and Boonyopakorn use it in developing regulatory lessons for Thailand. A Canadian Centre for Policy Alternatives guide brings the framework into public-facing discussion of national AI governance.

Other researchers have applied the comparison to Chinese newsroom AI, low-carbon energy systems, healthcare regulation, and advanced-AI governance.

The ICML position paper led by Francisco Eiras developed a benefit-risk framework for open-source generative AI and distinguished among different forms of openness.

Later researchers used that analysis in both policy and empirical work.

Paris, Moon, and Guo identified it at FAccT as one of the major frameworks for thinking about model openness. The Model Openness Framework and a TMLR consensus paper also drew on it.

More recent empirical research has tested those questions against actual practice. A large-scale study of artistic image-generation ecosystems used the paper to frame creator behavior across open models. The work has also entered research on cross-border business development in generative AI.

In 2016, Katherine Elkins and Jon Chun founded what Kenyon documents as the first Human-Centered AI curriculum and lab.

The curriculum treated disciplinary expertise as part of technical AI research rather than as an addition after model development.

The resulting work has entered scholarship on AI education and the role of the humanities in artificial intelligence. UNESCO’s Prospects quoted the program’s account of the relationship between digital humanities and AI. Klowden and Terence Tao also engaged its argument for retaining human judgment in questions of knowledge and value.

Jon designed the technical curriculum and courses as part of a research practice in programming, modeling, evaluation, and domain-specific inquiry.