evidoria

← Back to browse

Good practice Imported

Austria's AMS Job-Matching Algorithm (AMAS, 2018–2020) — A Cautionary Case of Gender Bias by Design

Austria · Vienna · See the Austria profile · See the Vienna profile

Evidence: Observational / pre–post Top 90% 19/100 · Ask Evidence Copilot about this practice

Austria's public employment service built a €240,000 statistical scoring tool (AMAS) that docked points for being female and for 'care obligations' (applied only to women), sorting jobseekers into support tiers. Austria's data-protection authority banned it in 2020 as unlawful pr

240,000 EUR
Development cost (Synthesis Forschung contract) (2018)
-0.14 model coefficient
Negative weight assigned for being female (2018-2020)
2.5 million EUR
Total project cost before termination (2018-2020)
Austria's AMS Job-Matching Algorithm (AMAS, 2018–2020) — A Cautionary Case of Gender Bias by Design

Details

Maturity
Discontinued
Promoter
Arbeitsmarktservice (AMS) — Austrian Public Employment Service
Period
2018–2020 (deployed and banned); 2025 (court review)
Region (NUTS)
AT13
Keywords
public employment services, algorithmic governance, data protection, gender equality

Context

Austria's public employment service (AMS) commissioned the research company Synthesis Forschung, for around €240,000, to build the 'Arbeitsmarktchancen-Assistenz-System' (AMAS) — an algorithm intended to score jobseekers' re-employment chances and sort them into three support groups: A (needs little help), B (retraining candidates) and C (low prospects, reduced support).

Objectives

The stated aim was to allocate AMS's employment-support resources more efficiently by predicting each jobseeker's chances of returning to work.

Activities

AMAS was deployed between 2018 and 2020, assigning jobseekers a score that determined which support tier they were placed in. Independent researchers (Allhutter et al., published in Frontiers in Big Data) analysed the published scoring model, and the digital-rights organisation AlgorithmWatch publicised the findings.

Results

The independent analysis found the model assigned a negative weight of −0.14 simply for being female, and included a 'childcare obligations' penalty variable that, by design, applied only to women. In August 2020, Austria's Data Protection Authority ordered the project halted, finding it lacked a legal basis and constituted prohibited automated individual profiling; AMS deleted the associated model and data, ending the roughly €2.5 million project. In September 2025, Austria's Federal Administrative Court ruled, on a narrower procedural question, that AMAS as designed would not have breached GDPR Article 22 — but since the underlying data had already been destroyed, the project could not be revived.

Conclusions

This is documented here as a cautionary rather than exemplary case: gender was encoded as a statistical penalty inside a public-sector allocation algorithm, and it took independent academic scrutiny and a data-protection authority's intervention to expose and stop it. It is relevant evidence for any AI-in-public-services practice that claims to be gender-neutral by design.

Implementation
Indicative cost
Medium (€50k–€500k) — Development contracted to Synthesis Forschung for roughly €240,000; total project cost before termination was approximately €2.5 million.
Time to results
Medium (1–3 years) — Deployed 2018-2020; halted by the Austrian Data Protection Authority in August 2020; the Federal Administrative Court ruled on a narrower procedural question in September 2025 but could not revive the project since the data had already been deleted.
Staffing & skills
Synthesis Forschung (external contractor building the scoring model), AMS internal project and data-protection team, Austria's Data Protection Authority (regulatory oversight)

Conditions for success

  • Independent bias and legal review of scoring variables before deployment
  • Exclusion of protected-characteristic proxies (gender, childcare obligations) from scoring models
  • Transparency of model variables and weights to allow external audit before rollout

Common failure modes

  • Gender was used as a direct negative-weighted variable in the scoring model
  • A 'childcare obligations' variable applied only to women, functioning as an indirect gender penalty
  • The system was deployed without an adequate legal basis for automated individual profiling, per the Austrian Data Protection Authority
  • Lack of transparency/external audit prior to rollout meant the bias was only caught after outside researchers examined the published model
  • Destruction of the underlying data on termination meant even a later, narrower favourable court ruling could not revive or correct the system

Where it fits

Governance type
national public employment service (algorithmic administrative decision-making)
Scale
national
Income level
high-income

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful