Mentored Projects

Timeline of firsts

Done at the frontier, as it moved

Selected projects, anchored to what existed — and what didn’t — when students did the work. Every name links to the full paper on Digital Kenyon.

• 2018
  Neural networks in a humanities classroom — before GPT-2 existedThe first AI for Humanity cohort trained recurrent networks to generate sheet music (Seth Colbert-Pollack), built convolutional style-transfer systems (Miles Shebar), applied deep reinforcement learning to trading (Tucker Bennett et al.), and ran NLP across thirty years of Ohio college newspapers (Shane Canfield) — the year before GPT-2 was announced.
• 2019
  Fine-tuning GPT-2 within months of its staged releaseOpenAI released the 345M GPT-2 model in May 2019 and withheld the full model until November. That same year, students were already fine-tuning 345M for creative work: Alasia Destine-DeFreece and collaborators asked whether GPT-2 could replace a Sex and the City writers’ room, and Olivia Kane framed “centaur screenwriting.” Others worked with StyleGANs (Sophie Dodd) and LSTM-generated 3D sculpture (Nick Downey).
• 2020
  The Gwern-cited cohort — and GPT-3 the year of its APIThree student papers on GPT-2 poetry and lyric generation — Jonah Zitelli on James Wright, Kaiya Case on John Donne, and Sophie Barrio on hit song lyrics — were cited and archived by Gwern Branwen, whose site remains a standing reference for early GPT-2 creative experimentation. GPT-3’s API opened in June 2020; by fall, Emmy Roday was studying GPT-3’s reimagining of Ginsberg’s “Howl,” and Rebecca Lawson used stylometry on AI-generated Nancy Drew, anticipating AI-text detection.
• 2021
  Systematic fine-tuning experimentsRebecca Lawson ran a hyperparameter study — temperature, epochs, corpus size — on GPT-2 screenplay generation, the kind of ablation then rare outside ML labs. Grant Holt asked whether a language model can write history; Teddy Kamin fine-tuned on Mike Royko’s journalism; Sarah Groustra trained GPT-2 on The Rocky Horror Picture Show to ask whether camp is computable.
• 2022
  Red-teaming ChatGPT within weeks of launchChatGPT launched November 30, 2022. Before the semester ended, Adam Blum had completed “Breaking ChatGPT with Dangerous Questions,” a systematic study of how the model prioritizes safety, context, and obedience — among the earliest jailbreak research anywhere. The same year, Fredrika Pfeiffer compared GPT-2 and GPT-3 on Das Kapital, Simon Hua treated prompt engineering as a research object before it had a literature, and Raya Kenney audited bias in image generators.
• 2023
  Auditing GPT-4 the spring it shippedGPT-4 arrived in March 2023; students audited it immediately. Abigail Foster tested whether GPT-4 could fool TurnItIn’s detector. Annalia Fiore evaluated GPT-4’s consistency in assessing — and self-assessing — literary texts, anticipating “LLM-as-judge.” Rishil Kondapaneni benchmarked ChatGPT against Google Bard; Dillon Cleary asked whether an LLM could stand in for an executive on an earnings call.
• 2024
  Studying real human-AI interaction at scaleHannah Sussman topic-modeled 3,275 real ChatGPT conversations as the first large conversation datasets became available. Richard Anthony Álvarez built a retrieval-augmented film recommendation system; Nava Bahrampour paired sentiment analysis of autism-diagnosis narratives with an audit of AI attitudes toward disability language.
• 2025
  Multi-agent systems, benchmarks, and domain fine-tunesGodwin Idowu tracked emotion, persuasion, and deception across multi-agent negotiations with valence-arousal-dominance modeling, and built a contamination-resistant math benchmark. Juliette Lowe fine-tuned BioLinkBERT for pediatric rheumatology. Marisol Hernandez Brito turned the program’s “shapes of stories” methods on AI’s own outputs; Parker Gibbons used multi-agent debate to automate MLB front-office decisions.
• 2026
  AI agent networksMuhammad Ibraheem Nadeem applied network analysis to Moltbook, an AI agent social network, identifying missing memory as a structural problem. Ayesha Aslam combined aspect-based sentiment analysis and behavioral analytics to study how AI influencers capture attention.