Jon A. Chun

Research focus

• Reasoning

  How frontier models form judgments and reason about ethics under ambiguity, and whether their decisions hold up when the wording changes but the logic does not.

• Behavior

  How intelligent systems act in real-world and adversarial conditions, studied through red-teaming, auditing and controlled perturbation rather than aggregate leaderboard scores.

• Interaction

  How AI systems engage with people and institutions, including strategic interaction between agents, pragmatic understanding across social hierarchies, and governance standards that connect evidence to policy.

Current roles

• Co-PI, MLA team, NIST AI consortium

  Lead AI safety researcher for the Modern Language Association, which represents 25,000 members, inside the US Department of Commerce consortium on trustworthy and safe AI. The MLA team, led with Katherine Elkins at Kenyon College, evaluates model capabilities with a focus on linguistic edge cases and ethical frameworks.

  2024 to present Consortium

• Co-PI and AI architect, Archival Intelligence

  Schmidt Sciences Humanities and AI Virtual Institute grant for "Archival Intelligence: Can AI Rescue Endangered Archives." An interdisciplinary team from Columbia University, Berklee College of Music and Louisiana State University works on multilingual Creole and Cajun newspapers and early New Orleans jazz materials captured with nothing more than smartphone photos.

  2025 to present HAVI Project site

• Co-founder, Kenyon College AI CoLab

  Founded with Katherine Elkins, the AI CoLab hosts what Kenyon documents as the first human-centered AI curriculum and lab. Students from nearly every department learn enough programming, machine learning and evaluation to bring their own disciplinary questions into original technical research.

  Founded 2016 Program Teaching

• Founder, SafeWeb (acquired by Symantec)

  Co-founded and led the world's largest online privacy service, which handled two billion transactions in eight months and took the first security investment made by In-Q-Tel. The company pivoted to enterprise SSL VPN and was acquired by Symantec for 26 million dollars. Two US patents came out of the work.

  Earlier career Career and patents

Selected figures

• 2 US patents on network security, from the SafeWeb and Symantec clientless VPN work
• 1 of 23 Schmidt Sciences HAVI teams selected worldwide from more than 600 applications
• 130,000+ Downloads of mentored student research across 4,700+ institutions in 198 countries
• 190+ Published undergraduate research projects on Digital Kenyon

A current finding

Featured software

Open-source research code, maintained on GitHub since 2019.

• SentimentArcs

  A large ensemble of dozens of sentiment models for analyzing emotion in text over time, and the technical basis for The Shapes of Stories (Cambridge University Press, 2022).

  Jupyter Notebook 43 stars

• MultiSentimentArcs

  The first fully open-source framework for diachronic multimodal sentiment analysis, comparing emotional trajectories across film dialogue and imagery.

  Python 6 stars

• AIStorySimilarity

  A benchmark for long-text story similarity built on narrative theory, with applications to search, intellectual property infringement and guided creativity.

  Python

• LLM Ethics Audit

  An ethics-based audit framework applied to the leading LLM chatbots, comparing the ethical frameworks they invoke and where those frameworks diverge.

  Audit framework 1 star

All software and every public repository

Selected recent papers

• Syntactic Framing Fragility: An Audit of Robustness in LLM Ethical Decisions

  arXiv preprint, December 2025

  Auditing 23 state-of-the-art models over 14 ethical scenarios and four controlled framings, a total of 39,975 decisions, finds widespread inconsistency. Many models reverse an ethical endorsement on syntactic polarity alone, and open-source models show over twice the fragility of commercial ones. Chain-of-thought reasoning substantially reduces it.

• AgenticSimLaw: A Juvenile Courtroom Multi-Agent Debate Simulation for Explainable High-Stakes Tabular Decision Making

  Accepted, AAAI-26 LaMAS, Singapore, January 2026

  Courtroom-style orchestration with explicit prosecutor, defense and judge roles, a seven-turn structured debate and private reasoning strategies. Structured multi-agent debate proves more stable and generalizable than single-agent reasoning across roughly 90 model and strategy combinations.

• AIStorySimilarity: Quantifying Story Similarity Using Narrative for Search, IP Infringement, and Guided Creativity

  ACL EMNLP/CoNLL 2024, Miami

• Near to Mid-term Risks and Opportunities of Open-Source Generative AI

  ICML 2024, Vienna, presented as an oral

Full publication list and work under review

News

• Feb 2026

  NPR and the Ohio Newsroom cover AI's potential to help preserve endangered archives, syndicated by Library Journal infoDOCKET. The Christian Science Monitor covers AI safety and the NIST consortium work the same month.

• Jan 2026

  AgenticSimLaw accepted at the AAAI-26 LaMAS workshop in Singapore. Papers under review at ICML, FAccT and other 2026 venues.

• Dec 2025

  Awarded a Schmidt Sciences Humanities and AI Virtual Institute grant for Archival Intelligence, one of 23 teams selected from more than 600 applications. Kenyon College reports the award.

• Nov 2025

  Forbes publishes "Where AI Meets the Humanities: Inside Kenyon College's Bold Experiment."

• Jun 2025

  Invited speaker at a Chronicle of Higher Education virtual forum.

• 2024

  Co-PI for the NIST AI Safety Institute Consortium on behalf of the MLA, and co-PI for the IBM and Notre Dame Tech Ethics Lab grant. First offering of Frontiers in Generative AI. Oral presentation at ICML 2024 in Vienna.