Generative AI and Cultural Creativity (2026-2027)
Background
Stories are central to how humans communicate, imagine and make meaning. Scholars across literature, cognitive science and anthropology often argue that compelling storytelling depends on surprise — the productive tension between what audiences expect and what a narrative actually delivers. But what counts as “surprising” changes over time, shaped by technologies such as print, film, digital media and now generative artificial intelligence.
Today’s generative AI systems are trained on massive corpora of human-authored stories, yet they tend to produce narratives that feel predictable or formulaic. They excel at linguistic fluency but struggle with genuine novelty and suspense. This limitation stems from how large language models (LLMs) learn: by reinforcing statistical patterns of what frequently occurs rather than exploring meaningful deviations.
This gap raises two intertwined questions:
- How can we model narrative surprise in a way that links humanistic theories of storytelling with computational methods?
- What does AI’s difficulty with surprise reveal about the nature of human creativity itself?
Answering these questions has implications for literary theory, cultural analytics and the design of future AI systems.
Project Description
This project will study how generative AI models represent narrative surprise and how their limitations can illuminate human creativity. Work will proceed in three phases:
Data collection and preparation
Team members will curate and annotate a corpus of 50-100 short narrative excerpts from public-domain or open-licensed sources. Students will identify moments of suspense or reversal and craft prompts that capture the narrative setup just before the “surprising” turn. Using these prompts, the team will generate story continuations from open-source AI models such as Llama, Mistral and Qwen (and, if feasible, select publicly available commercial models).
Analysis and evaluation
The team will write scripts to compare model-generated continuations with the original human-written stories. Quantitative metrics may include embedding-based diversity, entropy and distributional similarity. Qualitative analysis will evaluate narrative elements such as plot progression, pacing and surprise.
Small panels of peer readers and trained annotators will rate a subset of outputs for originality, coherence and engagement. Findings will test whether computational metrics meaningfully capture human judgments of creativity.
Synthesis and dissemination
The team will integrate quantitative and qualitative results into an annotated dataset and a reproducible analysis toolkit. Team members will build visualizations illustrating how different models handle novelty and predictability, interpret results through literary theory and prepare a short co-authored research paper.
Outputs will be shared through Duke’s Data+ and Bass Connections symposia, a public-facing GitHub repository and outreach to creative communities.
Throughout the year, students will gain training in text analysis, Python coding, statistical reasoning, literary interpretation and interdisciplinary collaboration.
Anticipated Outputs
- Annotated dataset of story prompts, human texts and AI-generated continuations
- Open-source analysis toolkit and reproducible Jupyter notebooks
- Visualizations and dashboards illustrating model behavior
- Co-authored paper and presentations at Duke symposia
- Public GitHub repository and explanatory website
- Student learning portfolios
- Foundational framework for future grants on AI and creativity
Student Opportunities
The team will include 3 graduate students and 8 undergraduate students from computer science, data science, English, literature, linguistics, visual and media studies, statistics and public policy.
Students will gain experience in:
- Text curation, annotation and corpus design
- Python programming and quantitative text analysis
- Designing and evaluating metrics for narrative novelty and surprise
- Literary interpretation and cultural theory
- Data visualization and public communication
- Mixed-methods research bridging humanities and computing
- Collaborative project management
Graduate students will mentor subteams, guide coding and annotation workflows and help connect humanistic theory to computational methods.
See the related Data+ project for Summer 2026; there is a separate application process for students who are interested in this optional component.
Timing
Summer 2026 – Summer 2027
Summer 2026 (optional):
- Curate narrative excerpts and generate AI continuations
- Build initial dataset and annotation schema
- Create prototype code notebooks and preliminary analyses
Fall 2026:
- Expand dataset and refine annotations
- Develop evaluation metrics and run reader-response tests
- Begin visualizing patterns in narrative surprise
Spring 2027:
- Complete analyses and visualizations
- Draft co-authored paper and launch public website
- Release open-source dataset and code
- Present findings at Duke events
- Plan outreach and future grant proposals
Summer 2027 (optional):
- Scale dataset, incorporate new genres and prepare follow-on publications
Crediting
Academic credit available for fall and spring semesters
See related Data+ summer project, Generative AI and Cultural Creativity (2026).