Report

Summary: Artificial Intelligence in Public Employment Services

Author: Sabrina Mühlbauer
Title: Artificial Intelligence in Public Employment Services: Machine Learning, Job Match Quality, and Fairness in Labour Market Matching
Institution: University of Regensburg, Germany
Dissertation: 2026, 122 pages. Read the complete dissertation (PDF).

1. Central argument

The dissertation examines how artificial intelligence (AI), particularly machine learning (ML), can improve the work of Public Employment Services (PES) in matching unemployed people with employment opportunities.

Its central argument is that the success of an employment service should not be measured simply by how quickly someone finds a job. It should also consider:

  1. The probability of successfully entering employment.

  2. The quality and stability of the resulting employment.

  3. The fairness of recommendations across different population groups.

The research uses extensive German administrative employment data to investigate these three dimensions.

The dissertation consists of three empirical studies, each addressing a different aspect of AI-supported employment matching.

2. The three principal findings

Study 1 – Machine learning improves occupational matching

The first study compares conventional statistical techniques with machine-learning approaches to predict which occupations an individual is likely to enter successfully.

Methods include:

  • Ordinary Least Squares (OLS).

  • Logistic regression.

  • Random Forest.

  • K-Nearest Neighbours.

Main finding: Random Forest substantially outperforms conventional econometric approaches in predicting occupational transitions.

The advantage of machine learning is its ability to identify nonlinear relationships and interactions among numerous variables, without requiring researchers to specify all those relationships in advance.

Policy implication: Employment counsellors could use predictive models to identify a broader range of realistic occupational opportunities for jobseekers, rather than relying exclusively on traditional occupational classifications or individual judgement.

Study 2 – Finding a job is not the same as finding a good job

The second study makes an important contribution by distinguishing the probability of obtaining employment from the quality of that employment.

It develops a job-matching index incorporating three elements:

  • Matching probability: likelihood of entering a particular occupation.

  • Employment stability: expected duration and sustainability of employment.

  • Expected wages: anticipated earnings associated with the occupation.

The resulting index provides a more comprehensive measure of employment success.

Main finding: Recommendations change substantially when employment stability and wages are considered alongside the probability of finding a job.

An occupation with a slightly lower probability of immediate employment may be preferable if it offers better wages and more stable employment prospects.

Policy implication: Employment services should not optimize their performance around rapid job placement alone. They should also promote sustainable employment and better longer-term labour market outcomes.

Study 3 – Algorithmic fairness and the risk of reproducing inequality

The third study examines one of the most important challenges associated with AI in public employment services: algorithmic discrimination.

Historical employment data contain existing inequalities, including occupational segregation by gender, nationality and age.

An AI system trained on these data may reproduce or even amplify those inequalities, despite having high predictive accuracy.

The study adopts an intersectional approach, examining combinations of demographic characteristics rather than treating each characteristic independently.

It distinguishes three dimensions of fairness:

  • Distributional fairness: whether recommendations reproduce or amplify existing occupational inequalities.

  • Predictive fairness: whether prediction errors and opportunities are distributed equitably across groups.

  • Operational fairness: whether residual risks remain in actual deployment.

Main finding: Improving fairness in one dimension does not necessarily improve fairness in others. Technical adjustments can reduce certain disparities but may also reduce predictive performance.

Importantly, technical measures alone cannot eliminate all risks.

Policy implication: AI-supported employment services require transparency, human oversight, institutional governance and meaningful participation by jobseekers.

3. Some important quantitative results

The dissertation provides useful numerical evidence, particularly concerning predictive performance and fairness.

Study 1 – Predictive accuracy

30%

Approximately lower out-of-sample classification error for Random Forest compared with OLS.

This is a substantial improvement in the ability to predict occupational transitions.

epub.uni-regensburg.de

 

Study 3 – Fairness versus predictive performance

Model Equal opportunity gap Balanced accuracy
No fairness adjustment 0.747 0.524
Moderate adjustment 0.598 0.472
Strong adjustment 0.184 0.381

Source: Table 3.2, page 90. Lower equal opportunity gaps indicate smaller predictive disparities.

epub.uni-regensburg.de

The third study reveals a particularly important trade-off:

  • A modest fairness adjustment has little effect on predictive accuracy but also produces limited fairness improvements.

  • A stronger adjustment substantially reduces predictive disparities, but at a significant cost in predictive accuracy.

  • The author selects an intermediate adjustment (λ = 0.001) as the preferred technical compromise, because it retains at least 95% of baseline predictive utility.

This is an important practical finding: maximizing fairness independently of the usefulness of recommendations is not necessarily a good solution.

The study also identifies residual disparities in occupations such as gastronomy, cooking, medicine, retail and advertising, even under the strongest fairness adjustment.

epub.uni-regensburg.de

 

4. An additional innovation: counterfactual averaging

The third study also explores a technique called counterfactual averaging.

The idea is to calculate predictions under alternative demographic identity assignments and average the resulting probabilities. This reduces the extent to which recommendations depend on demographic characteristics.

The results are encouraging:

  • Under the strongest fairness adjustment, the number of occupations identified as having particularly significant distributional distortions falls from five to two.

  • Average residual distributional divergence falls from 0.637 to 0.282.

However, this approach does not eliminate structural inequalities in the labour market. It reduces the sensitivity of the prediction mechanism to demographic identities.

This distinction is important: making an algorithm less sensitive to demographic characteristics does not automatically make the underlying labour market fairer.

epub.uni-regensburg.de

 

5. Implications for public employment services

The dissertation has several implications for the organization and delivery of employment services.

Traditional approach AI-supported approach
Counsellor identifies potential occupations Algorithm identifies and ranks potential occupations
Recommendations based largely on individual assessment Recommendations informed by large-scale employment histories
Emphasis on finding employment Emphasis on probability, stability and wages
Fairness assessed through general policies Fairness also assessed through measurable algorithmic outcomes
Counsellor makes recommendations Counsellor interprets and discusses algorithmic recommendations

The author does not advocate replacing employment counsellors with algorithms. Rather, AI should provide additional information and improve the quality of professional decisions.

The dissertation specifically emphasizes:

  • Explainability of recommendations.

  • Human oversight.

  • Jobseeker participation.

  • Monitoring of occupations where algorithmic distortions persist.

  • Institutional accountability for the use of AI.

These considerations are particularly relevant when employment services influence access to training, occupational mobility and employment opportunities.

6. Limitations and considerations

There are several important qualifications when interpreting the results.

1. German institutional context

The research uses German administrative employment data. The underlying database is a 10% sample of the Integrated Employment Biographies, covering employment histories and characteristics of workers subject to the German social security system.

epub.uni-regensburg.de

 

The results cannot automatically be transferred to Québec or Canada, where employment institutions, occupational classifications and administrative data differ.

2. Prediction is not causation

The models predict occupations people are likely to enter based on observed historical patterns. They do not demonstrate that following the recommendations will necessarily improve employment outcomes.

3. Historical inequalities remain embedded in the data

Even sophisticated fairness adjustments cannot fully eliminate the influence of historical occupational segregation and unequal opportunities.

4. Job quality is measured relatively narrowly

The second study focuses primarily on employment stability and wages. Other dimensions of job quality, such as working conditions, job satisfaction, career development, working hours and skills utilization, are not fully incorporated into the index.

5. Implementation remains to be demonstrated

The dissertation develops and evaluates predictive models using administrative data. It does not establish, through a real-world randomized implementation, the actual effects of deploying the complete system on unemployment duration, employment retention or earnings.

7. Overall assessment and relevance to Québec

I consider this dissertation particularly interesting for three reasons.

First: It challenges the conventional measurement of employment-service performance.

The distinction between getting someone into a job and helping someone obtain sustainable, better-paid employment is fundamental.

This has implications for how governments evaluate employment programs, establish performance indicators and allocate resources.

Second: It provides a concrete methodology for integrating AI into employment services.

Rather than discussing AI in general terms, the research demonstrates how administrative employment histories can be used to generate occupational recommendations, assess job quality and identify potential discrimination.

Third: It highlights the importance of institutional governance.

The research makes a convincing case that algorithmic performance cannot be separated from the responsibilities of public institutions. Human judgement, transparency and jobseeker participation remain essential.

My suggested applications to Québec

For Québec’s employment and workforce development policies, I would identify five possible applications:

1. Occupational mobility: Identify realistic alternative occupations for unemployed workers, including occupations outside their previous sector.

2. Training investments: Improve the identification of training opportunities by considering not only the probability of obtaining employment but also expected wages and employment stability.

3. Transferable skills: Identify occupations where workers could transition with relatively limited retraining.

4. Program evaluation: Develop performance indicators that measure employment retention, earnings progression and occupational mobility rather than simply the number of people placed in jobs.

5. AI governance: Establish mechanisms to detect whether algorithmic recommendations systematically limit opportunities for particular groups of workers.

8. The most important takeaway

The fundamental contribution of this dissertation is to move the discussion of AI in employment services beyond predictive accuracy.

A good employment recommendation system should identify not simply where someone is likely to find a job, but where that person has the best prospects of achieving sustainable and productive employment, without reinforcing existing inequalities.

For discussions concerning employment services, workforce development and the use of Employment Insurance Part II resources, I would consider this a particularly useful research contribution.

One particularly interesting avenue would be to extend the dissertation’s matching index to incorporate training needs and the potential return on investment in training. This would connect occupational recommendations directly to decisions about reskilling, upskilling and public investment in workforce development.

Source: Sabrina Mühlbauer, University of Regensburg, 2026 — Full dissertation (PDF, 122 pages).

Discussion

No comments yet.

Leave a Reply

Jobs – Offres d’emploi – US & Canada (Eng. & Fr.)

The Most Popular Job Search Tools

Even More Objectives Statements to customize

Cover Letters – Tools, Tips and Free Cover Letter Templates for Microsoft Office

Follow Job Market Monitor on WordPress.com

Enter your email address to follow this blog and receive notifications of new posts by email.

Follow Job Market Monitor via Twitter

Categories

Archives