Distributional Intersectional Fairness in AI-Supported Job Matching (IAB DP 06/2026)
A · Strong evidence Researchers at IAB, the German Federal Employment Agency's research institute, audit a real job-matching model on a 3.17 million-person administrative sample with a documented fairness method and per-occupation tables.
Summary based on the full text·Original: about 60 minutes to read
In short
IAB Discussion Paper 06/2026 by Sabrina Mühlbauer, Paula Ziethmann and Enzo Weber evaluates an AI occupation recommender being built for German public employment services. It measures fairness as the shape of outcome distributions across 60 combined gender-nationality-age groups, not as single-attribute gaps. The unconstrained model amplifies existing occupational sorting. Fairness penalties shrink aggregate disparity but cost accuracy and move distortions between occupations; a few occupations stay distorted whatever is done.
Key numbers
- 3,174,204individuals in the 10% random sample of German administrative labour-market records
- 141occupational classes the matching model recommends across
- 60intersectional groups formed from gender, nationality and age
- 0.637mean residual divergence in flagged occupations under strong regularisation (0.282 with counterfactual averaging)
- 0.636largest model-induced divergence shift (military occupations) in the unconstrained baseline
As reported by the source.
English·Web page·opens d.repec.org
What it reports
- The unconstrained model (gradient boosting, protected attributes held out of inputs) is "not distribution-preserving": a non-trivial share of occupations sit above the 45-degree line, meaning amplified concentration, and amplification exceeds attenuation in size.
- Largest amplification: military (shift +0.636 in Jensen-Shannon divergence), legislators and senior officials of interest groups (+0.362), trading occupations (+0.355).
- A fairness penalty on the training loss improves the equal opportunity gap monotonically, but only small penalties keep balanced and top-3 accuracy above 95% of baseline. The authors pick λ = 0.001; λ = 0.010 is treated as a stress test that degrades accuracy too far.
- An occupation-level flag (mean plus one standard deviation of the distortion) marks 20-22 occupations at low regularisation and 5 at the strongest setting.
- Counterfactual averaging, which averages predictions across perturbed identities, cuts flagged occupations from 5 to 2 and mean residual divergence from 0.637 to 0.282 at λ = 0.010, while keeping most predictive performance.
How they got it
Integrated Employment Biographies data rebuilt to mirror what caseworkers see in the placement system; the outcome is the occupation entered after unemployment. Observed group-by-occupation distributions are compared with model-implied ones using Jensen-Shannon divergence. Predictive fairness uses the equal opportunity gap; deployability uses a retained-utility rule on balanced and top-3 accuracy.
Limits
- Gender is binary in the data; disability and ethnicity beyond nationality are not observed.
- One class of fairness penalty and a few metrics are tested; other definitions could rank options differently.
- The flag threshold is relative to each model's own distortion spread, so flag counts are not comparable as absolute unfairness across settings.
- Counterfactual gains reflect reduced sensitivity in predictions, not removal of the labour-market inequality in the data.
Why it matters for companies
Any internal mobility, job-matching or candidate-ranking tool trained on past placements will copy and can sharpen past sorting, even with protected attributes removed. Single-attribute bias checks miss this; testing combined groups and per-role distributions catches it. The paper also gives a usable governance rule: set an explicit accuracy floor, pick the fairness setting that clears it, and keep a short list of flagged roles under human review. Expect residual risk, and plan monitoring rather than a one-off certification.
Summary based on the full text. Read the original for the full argument and data.Read the original on d.repec.org
- Tags
- Work and AI