New AI framework detects and governs bias in skills-based job matching
Two-stage system uses chatbots, agents, and social choice to ensure fair hiring.
Researchers Andrea Forster, Gregor Autischer, Dominik Kowald, and Simone Kopeinik from Know-Center Graz have published a paper proposing a two-stage framework for detecting and governing bias in skills-based job matching systems. The first stage addresses how candidates provide skills and preferences via chatbots, and how the system extracts and structures this information, identifying bias risks in profile formation. The second stage embeds this into a multistakeholder recommender system where candidate, company, and regulatory objectives are each represented by separate agents producing independent rankings. These rankings are combined through social choice-based aggregation into a single auditable recommendation.
The framework uses a distinction between hard constraints (requiring immediate correction) and soft constraints (logged for later decisions). Following an AI Act-aligned assessment methodology based on the Fraunhofer AI Assessment Catalog, the authors propose using distributional auditing and counterfactual testing to create a bias inventory. In Stage 1, soft constraints inform fairness thresholds for Stage 2. In Stage 2, fairness metrics crossing thresholds trigger an adapted recommendation process, while smaller deviations are logged as bias reports. The paper is accepted to the BIAS Conference 2026 in Leiden, Netherlands.
- Two-stage framework: skill extraction with chatbot bias detection, then multistakeholder recommendation via social choice aggregation.
- Uses hard constraints (must fix) and soft constraints (logged for later) to manage bias in real-time.
- Aligned with EU AI Act assessment methodology using distributional auditing and counterfactual testing.
Why It Matters
Provides a practical, auditable framework to ensure AI hiring systems are fair and compliant with regulations.