humancompatibleai/pareto

PARETO is a research dataset released alongside an academic paper studying whether AI systems can respond to politically sensitive questions in a way that feels fair to people across the political spectrum. The central question the paper explores is what political neutrality actually means for an AI: not simply avoiding strong opinions, but producing answers that people on different sides of political issues find roughly equally acceptable. The dataset contains responses from multiple AI models to a set of politically framed questions. Each response was shown to a large group of human survey participants (recruited through the Prolific research platform) who rated how much they approved of what the AI said. The survey also collected demographic information and written qualitative feedback. Prolific IDs were hashed to a consistent anonymized identifier to protect participant privacy while keeping responses linkable across the data files. The repository is organized into four main areas: raw AI model responses stored as CSV files, the same response pairings formatted as PNG images for display in the survey interface, the survey results (including numeric ratings and free-text comments from participants), and Jupyter notebooks that reproduce the charts and figures used in the published paper. The analysis code uses principal component analysis and correlation statistics to examine patterns in how different demographic groups rated AI responses. This is not a tool to run or install. It is a data archive for researchers who want to study AI political neutrality, reproduce the paper's findings, or use the survey methodology as a starting point for related work.