Michael Scott is Not a Juror

Limits of AI in Simulating Human Judgment

Authors

• Sean Harrington – Director of Technology and Innovation, University of Oklahoma College of Law
• Hayley Stillwell – Associate Professor of Law, University of Oklahoma College of Law

Funding

This research was supported by a grant from the Research Council of the University of Oklahoma Norman Campus.

About

This repository contains the data, code, and figures that underpin the paper’s analysis of how large language models compare with 1,200 human mock jurors across multiple criminal-law scenarios.

Directory Structure

data/       anonymised human responses and model outputs
scripts/    notebooks and utilities to reproduce results
paper/      camera-ready manuscript and figures
models/     fine-tuned Mistral-7B checkpoint (optional download)

Quick Start

# clone and set up
git clone https://github.com/your-org/juror-sim-paper.git
cd juror-sim-paper
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
# run analysis
jupyter lab               # open scripts/01_analysis.ipynb and Run All

License

Code: MIT Data: CC BY-NC 4.0 (non-commercial use, attribution required)

Pull requests and issues are welcome.