Beyond the Sidelines: Shaping AI Infrastructure with Labor Market Data

As social scientists, we focus heavily on how AI is changing the labor market. Historically, our discipline has not viewed itself as central to shaping these computational models. This perspective falls short. Both methodologically and substantively, data-producing institutions like the IAB have an urgent opportunity and responsibility, to actively intervene in AI development.

Generative AI pipelines are currently saturated with unscientific behavioral measurements. AI agents routinely deploy unstandardized, fixed-choice questions generated by the models themselves, while Reinforcement Learning from Human Feedback (RLHF) relies on non-probability annotator pools that introduce systemic selection bias. The fundamental machine learning challenge - determining whose values and life circumstances a model represents - is, at its core, a target-population estimation problem. Social science possesses the exact estimation machinery required to diagnose and repair these data structures.

This talk argues that as AI systems are deployed into high-risk labor market domains like matching and profiling, compliance with the EU AI Act demands rigorous auditability. This regulatory shift creates an immediate mandate for the methodologies we produce. In this presentation, we will map how labor market data can directly shape AI pipelines through representative benchmarks, validated instruments, and advanced error diagnosis - including adapting record-linkage traditions to measure what models get wrong about specific demographic groups. Furthermore, we will explore how sampling theory and classic survey methodology can govern the ways AI agents ask questions.

At the UN level, AI readiness for data collected by official institutions has become a central theme. I hope to discuss this element and its broader institutional implications following the presentation.

Date

Venue

Further information