Done at the frontier, as it moved
A decade of undergraduate AI research mentored by Jon Chun and Katherine Elkins in the human-centered AI curriculum they co-founded at Kenyon College in 2016 — humanities and social-science students doing the kind of original lab research that STEM students have always had.
Across 400+ original student research projects, one pattern repeats: a new capability ships, and the students Jon and Katherine Elkins mentor are auditing it, fine-tuning it, or building with it the same semester — often one to three years before the research community arrives at the same questions. The best 200+ are published on Digital Kenyon, where they have been downloaded 107,000+ times from 4,760 institutions across 198 countries (as of June 10, 2026).
In June 2026, The New Yorker's "Eight Predictions for the Future of Higher Education" named frontier-lab, project-based research a model higher education will adopt by 2035.
Jon Chun and Katherine Elkins have run that model since 2016.Anchored to what existed
- 2018Neural networks in a humanities classroom — before GPT-2 existed: the first AI for Humanity cohort trained recurrent networks for music generation, convolutional style transfer, and deep reinforcement learning, and ran NLP across thirty years of college newspapers.
- 2019Fine-tuning GPT-2 within months of its staged release — students fine-tuned the 345M model for creative writing and "centaur screenwriting" the same year it shipped.
- 2020Student papers on GPT-2 poetry were cited and archived by Gwern Branwen; by fall, students were already studying the freshly released GPT-3 API.
- 2021Systematic fine-tuning experiments — hyperparameter ablations on GPT-2 generation, the kind of study then rare outside ML labs.
- 2022Red-teaming ChatGPT within weeks of launch — a systematic study of safety, context, and obedience, among the earliest jailbreak research anywhere.
- 2023Auditing GPT-4 the spring it shipped — AI-detection evasion, literary self-assessment anticipating "LLM-as-judge," and cross-model benchmarking.
- 2024Studying real human-AI interaction at scale — topic-modeling thousands of real ChatGPT conversations; retrieval-augmented systems; disability-language audits.
- 2025Multi-agent systems, contamination-resistant benchmarks, and domain fine-tunes (e.g. BioLinkBERT for pediatric rheumatology).
- 2026AI agent networks — network analysis of AI agent social networks, and aspect-based sentiment analysis of AI influencers.
Representative projects
A representative selection across three Kenyon courses — Programming Humanity, AI for Humanity, and the Senior Seminar. Full papers for the best 200+ projects are on Digital Kenyon.
Generative AI & creative writing
- Can GPT-2 Replace a Sex and the City Writers' Room? — Alasia Destine-DeFreece et al., 2019 (first)
- Centaur Screenwriting with 345M GPT-2 — Olivia Kane, 2019 (first)
- 345M-GPT-2 After James Wright: Can AI Generate Convincing Contemporary Poetry? — Jonah Zitelli, 2020 (cited on gwern.net)
- The GPT-3 Re-Imagining of "Howl" — Emmy Roday, 2020 (first)
- Digitizing Camp: Training GPT-2 on The Rocky Horror Picture Show — Sarah Groustra, 2021
Model auditing, red-teaming & evaluation
- Breaking ChatGPT with Dangerous Questions: Safety, Context, and Obedience — Adam Blum, 2022 (weeks after launch)
- Can GPT-4 Fool TurnItIn? Testing the Limits of AI Detection — Abigail Foster, 2023 (first)
- Evaluating GPT-4's Consistency in Assessing — and Self-Assessing — Literary Texts — Annalia Fiore, 2023 (anticipates "LLM-as-judge")
- AI Proof Benchmarking: Mathematical Reasoning via Taylor Series Analysis — Godwin Idowu, 2025
Sentiment analysis & the shapes of stories
- Doubles and Reflections: Sentiment Analysis and Nabokov's Pale Fire — Catherine Perloff, 2019
- Many Stories, One Shape: Narrative Convergence in AI-Generated Fiction — Marisol Hernandez Brito, 2025
- Aspect-Based Sentiment Analysis with LLMs: Earnings Calls — Peyton Hodges, 2025
Multi-agent systems & the frontier
- Emotion, persuasion, and deception across multi-agent negotiations (valence-arousal-dominance modeling) — Godwin Idowu, 2025
- Multi-agent debate to automate MLB front-office decisions — Parker Gibbons, 2025
- Network analysis of an AI agent social network — Muhammad Ibraheem Nadeem, 2026
Who does this work
Browse all the work on Digital Kenyon → · See Katherine Elkins's project archive →