01
Reasoning
How AI systems make, explain, and justify judgments under ambiguity—including ethical reasoning, decision-making, and the relationship between an answer and its explanation.
AI evaluation · Computational research · Deployed systems
Jon A. Chun
Jon Chun is a researcher and builder working at the frontier of AI evaluation. He combines engineering, humanistic methods, and real-world deployment experience to test how intelligent systems reason, behave, and interact with people.
His work spans frontier-model evaluation, computational research into difficult human questions, the study of human judgment, and systems tested beyond the lab. Experience in privacy, security, and adversarial environments informs how he turns abstract claims about AI into concrete, testable questions.
Frontier evaluation
Co-PI, MLA team, NIST CAISI consortium
Human judgment, ethical reasoning, and model evaluation
Humanities and AI
Co-PI · Archival Intelligence
Schmidt Sciences Humanities and AI Virtual Institute · rescuing endangered cultural archives
Deployed systems
SafeWeb → Symantec
Privacy and security engineering under adversarial conditions
Research frame
What he works to understand
02
Behavior
How intelligent systems act in real-world, adversarial, and multi-agent conditions, with attention to robustness, agentic behavior, evaluation, and practical failure modes.
03
Interaction
How AI systems interact with people, institutions, culture, and context—from human-AI interaction and narrative to institutional use and deployment.
How he works
Methods that meet the problem
Humanistic methods are not decorative context. They help define what should be measured and show how ambiguous human concepts can be operationalized without reducing away the problem itself.
Engineering
Building systems, prototypes, technical infrastructure, and computational methods that make ideas testable.
Empirical evaluation
Turning claims about AI into benchmarks, measurements, experiments, and comparative analysis.
Humanistic methods
Using interpretation, construct definition, narrative analysis, context, and ambiguity to improve evaluation.
Real-world deployment
Testing systems against users, institutions, privacy and security constraints, incentives, misuse, and failure modes.
Selected work
Research, methods, and systems
Evaluation
Frontier AI Evaluation
Jon studies human judgment and ethical reasoning in frontier models and contributes to evaluation and standards work as Co-PI representing the Modern Language Association in the NIST CAISI consortium. His research develops concrete ways to compare what models decide, how firmly they decide it, and how they explain those judgments. Explore the evaluation research →
Computation + interpretation
Computational Human Questions
SentimentArcs, narrative modeling, and explainable-AI research reflect a longer project: turning complex human constructs into computationally testable forms while preserving their ambiguity and context. The work joins scalable analysis with interpretable measures. See SentimentArcs →
Systems in use
Building in Adversarial Environments
Building privacy and security systems at SafeWeb and Symantec taught Jon to test technological claims against actual users, incentives, misuse, and failure modes. That experience now informs his approach to AI evaluation. Read the building story →
Selected firsts
Methods and programs that established new ground
- 2016The world’s first Human-Centered AI curriculum and lab; Jon designed the technical curriculum and courses and developed the broader educational model with Katherine Elkins.
- 2019–21The first rigorous methodology for narrative sentiment analysis, the contribution of “middle reading,” and SentimentArcs’ novel ensemble method.
- 2020–23The first writer’s Turing test of an LLM and the first application of explainable AI to narrative.
- 2024The first ethics-based audit of moral reasoning in deployed LLMs; the first systematic post–EU AI Act comparison of EU, China, and US AI regulation; and the first multimodal method for measuring long-form film sentiment-arc coherence across modalities.
Current finding
A result that changes the comparison
110–300×
Across healthcare, law, and finance scenarios, instruction-tuned language models were 110–300× more resistant to narrative manipulation than people. The finding shifts attention from whether models can be influenced in principle to how model and human vulnerabilities differ in practice. Read the research →
Selected recognition & uptake
Where the work travels
Research on narrative AI, model openness, ethical reasoning, and human-centered AI has been taken up across computer science, digital humanities, governance, and education.
Selected scholarly reception and public discussionScholarly reception → · Grants and recognition → · Press coverage →
Pages
Research and background
AI evaluation, human judgment, narrative modeling, publications, and current findings.
SafeWeb, privacy and security infrastructure, patents, and deployed AI systems.
Archival Intelligence, NIST CAISI, research partners, students, and institutions working on shared problems.
The longer biography, background, teaching, speaking, and full professional record.
Get in Touch
Start a conversation
For research, collaboration, speaking, media, and professional inquiries.