From Survey Data to AI Alignment: Bridging Social Science and Machine Learning
This talk is organised in cooperation with the Academy of Sociology.
Abstract:
Large language models increasingly shape value-laden decisions, yet systematically fail to represent diverse human values. Current alignment approaches rely on dominant online discourses and unrepresentative feedback at single points in time. This talk introduces adaptive alignment: a framework that grounds AI in decades of population-representative social science data. With this talk I am to discuss with MZES the idea of living benchmarks that evolve with society, privacy-preserving methods for sensitive attitudinal data, and pluralistic approaches that learn from disagreement. Such efforts can bridge the gap between survey methodology and machine learning, and can create AI systems that are legitimate, representative, and responsive to Europe's diverse cultural landscape.
