What Is the Algorithm?

AMAS didn't decide who got a job. It decided who Austria's public employment service would spend more time and money helping - by running each jobseeker's history through a statistical model and sorting them into three tiers.

The problem AMS said it was solving

Austria's Public Employment Service (AMS) has a limited budget for job-search assistance - things like retraining courses, coaching, and wage subsidies - spread across a large number of unemployed people. Starting in 2016, AMS ran a program to study which jobseekers actually benefited most from that kind of help, with the stated goal of directing resources more efficiently rather than spreading them evenly. (source: AlgorithmWatch)

How AMAS worked

By 2019, AMS had commissioned an external contractor, Synthesis Forschung, to build a statistical model - at a cost of 240,000 euros - based on regressions run against past employment records. The model was designed to predict each jobseeker's "labor market chances" and automatically assign a score. Depending on that score, a person would land in one of three groups: Group A, for people judged to need no help finding a new job; Group B, for people who might benefit from retraining or other support; and Group C, for people deemed to have poor prospects over the next two years, who would receive less intensive - and less expensive - help from AMS. (source: AlgorithmWatch)

What made it controversial

Internal AMS documentation reviewed by AlgorithmWatch in 2019 showed the model assigned a negative weight to being a woman, to having a disability, and to being over 30 - and that women with children were penalized while men with children were not. In other words, factors that have nothing to do with a person's individual effort or qualifications could lower their score and, in turn, the level of help AMS offered them. AMS's director, Johannes Kopf, responded to critics at the time by noting the agency was committed to spending half its resources on women and that women were underrepresented in the lowest-scoring group - without denying that the model itself treated those factors as negative. (source: AlgorithmWatch)

GDPR Article 22, explained

Despite that discrimination-shaped controversy, the legal fight that actually played out in Austria's courts wasn't fought under an anti-discrimination law - it was fought under Europe's data protection law, the GDPR, and specifically Article 22. That article restricts decisions "based solely on automated processing" - meaning no real human judgment involved - when the decision has a legal or similarly significant effect on a person. It doesn't ban automated scoring outright; it requires that a human meaningfully review and be able to override the result, not just rubber-stamp it. That's why Austria's courts spent years examining a narrower, more technical question than "is this scoring system biased" - they asked whether AMS advisors genuinely had the power to disagree with AMAS's score in practice, or whether the score effectively made the decision for them. See Why It's Significant for how that question was ultimately answered - twice, with different results five years apart. (source: The International, ppc.land)

A note on sourcing for this page: the underlying DSB and court decisions are German-language legal documents held on Austria's public legal database, ris.bka.gv.at. We located the official records there but could not reliably load their full text through our research tools during this case's build. Where that happened, we relied on English-language secondary reporting from specialized legal outlets (dataprotect.at, ppc.land) and news coverage (The International, ORF) that had read those documents directly, and we've named each one on every page.