Frequently asked questions

Who is Jon Chun?

Jon A. Chun is an AI researcher and educator and a technology entrepreneur. He co-leads the team representing the Modern Language Association at the NIST Center for AI Standards and Innovation (CAISI), is co-PI of the Schmidt Sciences HAVI project Archival Intelligence, and co-founded the world's first Human-Centered AI curriculum and lab at Kenyon College with Katherine Elkins (2016). Earlier he co-founded and led SafeWeb, acquired by Symantec in 2003.

What is Human-Centered AI as Jon Chun practices it, and how is it different from UX design, AI ethics, or digital humanities?

It uses state-of-the-art AI and real engineering to turn the oldest human questions into quantifiable, testable, metrics-based models. It builds, measures, and governs. It is distinct from human-centered UI/UX design (surveys, usability), from non-technical AI-ethics or STS critique, from low-code digital humanities, and from siloed single-discipline academic research. What is distinctive is the pairing of frontier-model engineering with humane breadth, through radical collaboration across disciplines, industry, government, and non-profits. More on Human-Centered AI →

Who founded human-centered AI, and how does Kenyon's approach differ from Stanford's?

Katherine Elkins and Jon Chun founded the world's first human-centered AI curriculum and lab at Kenyon College in 2016. Stanford's Institute for Human-Centered AI was founded in 2019 and approaches AI largely from computer science; the Kenyon approach enters from the social sciences and humanities, treating them both as a lens on AI and as fields AI can advance. See the full definition →

What is the 110–300× robustness paradox?

In studies of syntactic framing fragility, instruction-tuned large language models proved 110–300× more resistant to narrative manipulation than people, measured across healthcare, law, and finance vignettes. On this axis the models are far harder to fool than humans, a counterintuitive and measurable result about where LLM risk does and does not lie. See the research →

What is the confidence-scoring method for auditing language models?

Introduced in "Informed AI Regulation," it measures how firmly a model commits to a moral judgment versus hesitates, giving a way to compare normative certainty across models. It has since been applied across 1,613 social-decision scenarios (COLING 2025) and included among 69 foundational works in the AAAI 2026 "Beyond Verdicts" survey. It is co-authored with Katherine Elkins.

What is SentimentArcs?

SentimentArcs (arXiv:2110.09454, 2021) is a self-supervised ensemble method for diachronic sentiment analysis of narrative: dozens of sentiment models are passed over the same text and weighed against one another. It also introduced dynamic time warping to diachronic sentiment analysis of narrative, so arcs of unequal length can be compared by distance. The paper reports a negative result: state-of-the-art transformers can struggle to find narrative arcs, a limit of scale rather than of tuning. Its open-source implementation, maintained since 2019, is the technical foundation of The Shapes of Stories (Cambridge University Press, 2022).

What is MultiSentimentArcs?

MultiSentimentArcs (Frontiers in Computer Science, 2024) extends narrative trajectory analysis from text to film. It is a multimodal method (text and image, not multiple languages) that measures whether the sentiment arc recovered from a film's dialogue and the arc recovered from its images are in fact the same arc, making cross-modal coherence in long-form narrative measurable rather than asserted.

What did the eXplainable AI with GPT-4 paper contribute?

It brought explainable AI to narrative, a domain where XAI had been largely confined to image classifiers, and supplied two things the field reuses: a sentence-level story-trajectory methodology, and the agglomeration and comparability solutions that let narratives of unequal length be compared on shared coordinates. Co-authored with Katherine Elkins in the International Journal of Digital Humanities (2023).

What is the through-line of the computational narrative research?

Building the instruments that make computational reading of narrative testable. "Middle reading" (2019) named a position between distant and close reading; SentimentArcs (2021) replaced the single trusted model with an ensemble; eXplainable AI with GPT-4 (2023) made story trajectories explainable and made narratives of unequal length comparable; MultiSentimentArcs (2024) carried the method across modalities, from text to film. Katherine Elkins's complementary program asks a different question: what stays the same and what changes when stories travel across languages, cultures, and time. See the novelty boundary →

What does dynamic time warping do for sentiment arcs?

Two novels rarely tell the same story at the same pace or the same length. Dynamic time warping (DTW) computes the distance between two whole arcs while absorbing the temporal shifts and stretches between them, so arcs that share a shape but not a tempo are recognized as similar. SentimentArcs (2021) introduced DTW to diachronic sentiment analysis, pairing it with LTTB downsampling (which reduces every arc to a common number of points while preserving its peaks, valleys, and endpoints) and using the resulting distance matrix to cluster arcs hierarchically. Across corpora running from roughly 1,400 to 13,000 data points per novel, this is what makes arcs of unequal length comparable by distance. See the novelty boundary → Narrative trajectories & DTW → Translation & cultural transmission (Elkins) →

What was SafeWeb, and how does it connect to the AI-safety work?

SafeWeb was an internet-privacy company Jon co-founded in 2000 and led as CEO. It ran one of the largest web-anonymization services of its era, received the first security investment from In-Q-Tel (the CIA-affiliated venture fund), and was acquired by Symantec in 2003 for $26 million, producing two US patents on early SSL/clientless VPN appliances. The through-line to today's NIST CAISI work is a builder's habit of testing public claims against adversarial use and failure modes. SafeWeb on Wikipedia →

What is the NIST CAISI / MLA role?

Jon co-leads the only humanities-led team in the federal AI-safety consortium, representing the 25,000-member Modern Language Association at the NIST Center for AI Standards and Innovation (CAISI). The team works on LLM evaluation, red-teaming, and ethics-based auditing.

What is Archival Intelligence / Schmidt Sciences HAVI?

Archival Intelligence is a Schmidt Sciences Humanities and AI Virtual Institute (HAVI) project on which Jon is co-PI, 1 of 23 teams selected worldwide from 600+ applications. It builds free, open AI tools to rescue New Orleans' endangered Creole and Cajun multilingual newspapers and early jazz artifacts, and to confront cultural flattening in AI models. More →

Where can collaborators, journalists, and students start?

Journalists and the public: see Press and the featured findings. Academics and grant officers: see Research and Reception. Industry and investors: see Building. Students: see Teaching and the AI CoLab. Or get in touch directly.