Reception & influence
How research Jon Chun has co-authored, on language-model evaluation, AI regulation, and computational narrative, is taken up across fields. Named citers and structural adoption, not citation counts.
The work is now part of the reference layer other researchers build on. Its three-regime map is standard shorthand for how the EU, China, and the United States diverge; its ethics-based audit is a reference method at the AAAI/ACM Conference on AI, Ethics, and Society; and its narrative methods are a scaffold other teams build on. Uptake runs from Terence Tao to PNAS Nexus and into chemistry, medicine, and the technical machine-learning literature. Most of this work is co-authored with Katherine Elkins, whose reception page documents the same corpus in further detail.
Evaluating and governing frontier models
Informed AI Regulation: Comparing the Ethical Frameworks of Leading LLM Chatbots
The first ethics-based audit of moral reasoning in deployed LLMs, the paper that operationalized ethical evaluation as a systematic audit, and now the anchor of a strand of AI-safety evaluation work. Liu and colleagues' INVP framework (COLING 2025) directly adopts its confidence-scoring methodology for evaluating LLM value priorities; Sowmya and Vasudeva (IEEE Access, 2026) replicate the eight-model audit design outright. At AIES, Jain, Calacci and Wilson cite its central finding that LLM ethical reasoning shows a clear bias toward societal and cultural norms, and the Edinburgh LLM Ethics Whitepaper names it a representative methodology for probing ethical values through prompting.
Comparative Global AI Regulation: Policy Perspectives from the EU, China, and the US
Comparative Global AI Regulation is a standard reference for the three-regime (EU–China–US) comparison of AI governance. The first systematic EU–China–US regulatory comparison after the EU AI Act's passage, its three-regime map (horizontal risk-based, market-driven, state-led vertical) has become standard shorthand in the citing literature. Floridi and Ascani cite it in Minds and Machines on AI for legislative innovation in the Italian Parliament; Olugbade engages it across a Global Public Policy and Governance argument for a global governance mechanism; and its regulatory-sandbox recommendation is adopted in Lu and Tie's ASEAN–EU comparison. Uptake runs through Communications of the ACM, JCMS on EU digital sovereignty, Information Fusion on trustworthy AI in healthcare, and FAccT red-teaming of the AI Act; Vaughan C. Turekian draws on the three-regime map in PNAS Nexus. The map has even migrated into energy-systems engineering, drawn on in a comparative review of low-carbon power systems, evidence of cross-field travel.
Near to Mid-term Risks and Opportunities of Open-Source Generative AI
This position paper set the benefit–risk terms of the open-model governance debate. Taeihagh's "Governance of Generative AI" (Policy and Society, 2025) carries its access-spectrum analysis; Liesenfeld and Dingemanse build on it in their FAccT "open-washing" critique of the EU AI Act; and Casper, O'Brien, Longpre and colleagues cite it in their TMLR consensus paper on open-weight model risk, authored by safety researchers across MIT CSAIL, Stanford, Mila, Carnegie Mellon, Princeton, the UK AI Security Institute, and Hugging Face.
Measuring the shape of stories
Can GPT-3 Pass a Writer's Turing Test?
Can GPT-3 Pass a Writer's Turing Test? is the most-cited article in the history of the Journal of Cultural Analytics. The first writer's Turing test of a large language model, published within months of GPT-3's release, and cited within the year by Floridi and Chiriatti for the claim that "GPT-3 writes better than many people." The paper now sits in the reference layer of the LLM literature, and both sides of the capability debate quote it. It appears in PNAS (Mei and colleagues' behavioral Turing test), Science Advances (Spitale and colleagues on GPT-3 disinformation), the most-cited paper on LLMs in education (Kasneci and colleagues), and NeurIPS, EMNLP and NAACL work on impersonation and narrative bias. In literary studies, N. Katherine Hayles engages it in Bacteria to AI (2025) and Martin Paul Eve cites it across two books.
The Shapes of Stories: Sentiment Analysis for Narrative
The Shapes of Stories is the most-cited Element in Cambridge's Digital Literary Studies series. The first methodology for sentiment analysis of narrative: an ensemble approach benchmarking more than three dozen models and showing how smoothing, model selection, and interpretation determine what an emotional arc reveals. The sciences now test its claims at scale: Knight and Rocklage's Science Advances study of narrative reversals and He, Breithaupt, Kübler and Hills's study in Scientific Reports both ground their designs in emotional-arc analysis. The method has traveled to customer-journey research at Harvard Business School, game-feel design in IEEE Transactions on Games, and sustainability science. It has also traveled into non-Western traditions: in Humanities and Social Sciences Communications (Nature portfolio), Belhaouari and colleagues apply its emotional-arc approach to the Qur'anic Surah Yusuf, reading it against Western linear narrative.
Can Sentiment Analysis Reveal Structure in a Plotless Novel? (and "middle reading")
This paper introduced "middle reading" and was the first to test whether sentiment methods survive nonlinear narrative, the question computational literary studies still benchmarks against. The Aarhus group cites "Elkins and Chun 2019" across at least six papers on Hemingway, Danish literature, and literary-quality prediction as the standing reference for that question; Sui, Hamilton and colleagues call nonlinearity "a hard problem for narratology, by both computational (Elkins and Chun, 2019) and traditional approaches."
SentimentArcs: Self-Supervised Sentiment Analysis of Narrative Time Series
The ensemble method the arcs are built on: dozens of sentiment models passed over the same narrative and weighed against one another, together with the negative result that state-of-the-art transformers can struggle to find narrative arcs, a limit of scale rather than of tuning. It is also where dynamic time warping enters diachronic sentiment analysis: DTW computes the distance between two whole arcs while absorbing the temporal shifts and stretches that separate otherwise similar story shapes, and those distances drive the hierarchical clustering of arcs. That is what makes narratives of unequal length comparable by distance. Its uptake is chiefly as infrastructure. The open-source implementation, maintained since 2019, is the technical foundation of The Shapes of Stories (Cambridge University Press, 2022), the machinery that book's method runs on, and remains Jon's most-starred repository. Researchers have applied the toolkit to novels, fan fiction, games, film and television scripts, end-of-life medical narratives, and economic-crisis discourse.
eXplainable AI with GPT-4 for Story Analysis
The paper that brought explainable AI to narrative, a domain where XAI had been largely confined to image classifiers. It supplies the sentence-level story-trajectory methodology together with the agglomeration and comparability solutions that let narratives of unequal length be compared on shared coordinates. The journal's special issue on Reproducibility and Explainability names it as proposing "an advanced XAI approach and workflow for LLM-based research." Cugurullo and Xu carry it into political theory in Policy and Society, citing it as the authority on LLMs' opaque epistemology; surveys in Discover Applied Sciences and IEEE Access catalogue it as the exemplar of GPT-4 story analysis.
MultiSentimentArcs: Measuring Coherence in Multimodal Sentiment Analysis
The first multimodal method for measuring whether a long-form film's sentiment arcs cohere across modalities: whether the arc its dialogue traces and the arc its images trace are in fact the same arc. Where the earlier work made narrative trajectory measurable in text, this extends the ensemble approach to film and makes cross-modal coherence measurable rather than asserted, with a human-in-the-loop design in place of a single opaque model. Published open-access in Frontiers in Computer Science, it is the most recent of the narrative instruments and the one on which the reception record is still being written.
A founding statement
The Crisis of Artificial Intelligence: A New Digital Humanities Curriculum for Human-Centred AI
The Crisis of AI is among the most-read and most-cited articles in the International Journal of Humanities and Arts Computing. A founding statement of human-centered AI in the digital humanities — the first peer-reviewed account of what the field is for and how the humanities should meet AI. Terence Tao (with Tanya Klowden) draws on it for the observation that the tasks AI automates "did not require an understanding of more philosophical aspects of a profession, such as the nature of knowledge, beauty, meaning." UNESCO's Prospects quotes its diagnosis of the "AI crisis" in the digital humanities; engagement runs through chemistry (Journal of Chemical Education) and medical education (Informatics in Medicine Unlocked). Dong and colleagues' systematic review in Digital Scholarship in the Humanities (Oxford) treats it as a field reference, and Omri Asscher's AI & Society argument for a humanities-based study of AI builds on it alongside two of the authors' other papers.
See the full research record → · Grants & honors → · Google Scholar →
Who builds on the work, and where
The work bridges communities that rarely cite one another, supplying methods, frameworks, and reference points that each field takes up in its own terms.
The fields that take up the work
Where the work is read
Countries where indexed research cites the work, drawn from structured affiliation data. A coverage view: books, chapters, and non-English venues are under-represented, so the actual reach is wider than shown. Hover a marker for the country name.
In what fields is the work cited?
Literature and philosophy, the computational and digital humanities, AI safety and governance, AI authorship and the university, with cross-citations in the social sciences, medicine, and engineering.
What is the work best known for methodologically?
Sentiment-analysis methods for measuring the emotional shape of narrative, the first writer's Turing test of a large language model, and a comparative framework for AI regulation across the EU, China, and the US.
Where is the work read?
Across more than seventy countries: in North America and Europe and through South and Southeast Asia, the Middle East, North Africa, Latin America, and Sub-Saharan Africa.